a survey of selected indoor positioning methods …...reliable indoor positioning. among...

Tampere University of Technology

A Survey of Selected Indoor Positioning Methods for Smartphones

CitationDavidson, P., & Piche, R. (2017). A Survey of Selected Indoor Positioning Methods for Smartphones. IEEECommunications Surveys and Tutorials, 19(2), 1347-1370. https://doi.org/10.1109/COMST.2016.2637663

Year2017

VersionPeer reviewed version (post-print)

Link to publicationTUTCRIS Portal (http://www.tut.fi/tutcris)

Published inIEEE Communications Surveys and Tutorials

DOI10.1109/COMST.2016.2637663

Copyright © 2016 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all otheruses, in any current or future media, including reprinting/republishing this material for advertising or promotionalpurposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of anycopyrighted component of this work in other works.

Take down policyIf you believe that this document breaches copyright, please contact [email protected], and we will remove accessto the work immediately and investigate your claim.

Download date:31.05.2020

https://tutcris.tut.fi/portal/en/persons/pavel-davidson(39f319b3-83cf-44f0-aec2-db101c0bbc22).html

https://tutcris.tut.fi/portal/en/persons/robert-piche(aa9ffa8c-7051-44f3-befd-7d3250af402d).html

https://tutcris.tut.fi/portal/en/publications/a-survey-of-selected-indoor-positioning-methods-for-smartphones(c3573175-1bb1-481c-ac13-fae62e72ae72).html

https://doi.org/10.1109/COMST.2016.2637663

https://tutcris.tut.fi/portal/en/publications/a-survey-of-selected-indoor-positioning-methods-for-smartphones(c3573175-1bb1-481c-ac13-fae62e72ae72).html

https://doi.org/10.1109/COMST.2016.2637663

1

A Survey of Selected Indoor Positioning Methodsfor SmartphonesPavel Davidson and Robert Piché

Department of Automation Science and Engineering, Tampere University of Technology, Finland

Abstract—This paper provides an overview of the most signif-icant existing methods for indoor positioning on a contemporarysmartphone. The approaches include Wi-Fi and Bluetooth basedpositioning, magnetic field fingerprinting, map aided navigationusing building floor plans, and aiding from self-contained sensors.Wi-Fi and Bluetooth based positioning methods considered inthis survey are fingerprint approaches that determine a user’sposition using a database of radio signal strength measurementsthat were collected earlier at known locations. Magnetic fieldfingerprinting can be used in an information fusion algorithmto improve positioning. The map-matching algorithms includeapplication of wall constraints, topological indoor maps, andbuilding geometry for heading correction. Finally, methods forstep counting, step length and direction estimation, orientationtracking, motion classification, transit mode detection, and floorchange detection in multi-storey buildings are discussed.

Index Terms—Indoor positioning, WLAN, BLE, magnetome-ter, fingerprinting, map-matching, IMU.

I. INTRODUCTION

Since every smartphone is equipped with GNSS or AssistedGNSS receiver, street navigation with smartphones is ubiqui-tous. However, there is no single technology that can providereliable indoor positioning. Among technologies suitable forindoor positioning that potentially can be implemented onsmartphones are:

• Use of radio frequency (RF) signals, either those alreadypresent such as wireless local area networks (WLAN),or cellular, or those generated by new and dedicatedinfrastructure (RFID/NFC, Bluetooth).

• Self-contained sensors: tri-axis accelerometer, tri-axis gy-roscope, tri-axis magnetometer, barometer

• Indoor map (building floor plan)• Magnetic fileld fingerprinting

None of these technologies is a complete solution to allindoor positioning needs on mass-market devices and theiradvantages and disadvantages depend on the nature of eachtechnology. For example, UWB- and RFID- based solutionsrequire deployment of new infrastructure and addition of newhardware components to smartphones. These additional costscurrently limit use of tag-based solutions to Bluetooth LowEnergy (BLE).

Typical tasks for indoor navigation with smartphone includelocating people and places in public buildings like universities,malls and airports, at sport events, conventions, etc., keepingtrack of children and elderly people, and automatically taggingposts and pictures. To meet these requirements, positioningaccuracy in a typical office building has to be at the room level

!

Magnetic!field!map!

Fusion!algorithm!

Map5matching!Indoor!map!!!

Coordinates!Location!on!the!map!

Smartphone!sensors!!Accelerometers!

Gyros! Compass!

Basic!positioning!module!

Magnetic!field!fingerprinting!

BLE!based!positioning!

!

WLAN!based!positioning!

!

Fig. 1. Major components of a smartphone indoor positioning system.

and the floor has to be correctly identified. In the case of largeopen spaces such as airport and shopping malls an accuracyof several meters is enough to establish visual contact andfind the required place, for example, an airport gate. For someapplications like store navigation or a guided tour of museumsbetter positioning accuracy is required. It can be achieved byBLE based positioning when the beacons are densely placedin the required areas.

Currently most of smartphones use WLAN based indoorpositioning that usually satisfies the position accuracy re-quirements in areas rich with WLAN routers. However, theaccuracy significantly depends on the density of WLANrouters and on the quality of the radio map which usuallydeteriorates over time and has to be updated on a regularbasis. Improvement of WLAN/BLE based positioning can beachieved if additional sensors and sources of information areused, for example self-contained sensors that improve accuracyand reliability between WLAN based position updates andduring gaps in WLAN/BLE coverage. Other commonly usedmethods to improve positioning are map-matching and mag-netic field fingerprinting. They can be fused with WLAN/BLEand sensors to improve position accuracy. Magnetic fieldbased positioning is conceptually similar to terrain navigationsystems in military aircraft and submarines, in which altitudeor depth are map-matched to estimate the position; in magneticindoor positioning the features are the local distortions inmagnetic field caused mainly by steel in the building structure.

The major components and data flow in a smartphone posi-tioning system are shown in Fig. 1. These technologies meet

2

the requirements for mass market deployment on smartphones,however not necessarily all of these components have to bepresent in the positioning system. Possible combinations oftechnologies for indoor navigation include:

• WLAN (map is used for display purpose only)• WLAN and sensors• WLAN and indoor map for map-matching• WLAN, map-matching and sensors

Magnetic field fingerprinting and BLE based positioning canbe added to any of these combinations. This paper describesthe technologies and position-estimation methods of thesetechnology components.

The paper is organized as follows. Section II considersWLAN and BLE based positioning systems. Magnetic fieldfingerprinting is explained in Section III. Map-matching ap-proaches suitable for implementation in indoor environmentare presented in Section IV. Section V surveys the use ofsmartphone sensors to improve positioning and to bridge gapsin WLAN and Bluetooth coverage. Finally, Section VI sum-marizes this survey of approaches to indoor positioning usingsmartphones and gives recommendations for future researchin this area.

II. WLAN AND BLE BASED POSITIONING

WLAN access points (AP) broadcast beacon frames, whichinclude the AP’s media access control (MAC) address, typi-cally every 100 ms to announce their presence in a certainarea [1]. The mobile nodes receive these signals and canidentify the AP according to its MAC address. AlthoughWLAN was not designed for positioning, it can be usedto estimate user location by exploiting the received signalstrength (RSS) measurements. This technology is very attrac-tive for positioning because WLAN APs are readily availablein most buildings and every smartphone has WLAN connectiv-ity. However, WLAN positioning requires the generation andmaintenance of a radio map, which is a time-consuming andlabor-intensive task. Using crowdsourcing for the radio map isdifficult because reference position data is usually not availableindoors, and so collection of positioning data cannot be done inthe background on user devices. Smartphone implementationof WLAN based positioning is usually based on the RSSmethods. Since the measured RSS data is very noisy onlytwo-stage approaches that include the preliminary stage ofthe training data collection can provide required localizationaccuracy.

This section overviews some established fingerprinting al-gorithms such as nearest neighbour (NN), k-nearest neighbour(KNN), weighted k-nearest neighbour (WKNN), coverage areaand path-loss (PL) which are the most common approaches inRSS based positioning on smartphones (Fig. 2) as well asrecently emerging techniques such as state vector machines(SVM), multiple weighted decision trees, compressive sensing,channel state information (CSI) and deep learning.

A. RSS based positioning

A basic approach to positioning is trilateration (Fig. 3).This technique estimates the location of a mobile node by

Radiomap(RSSfingerprints)

RSSmeasurements

Trilateration

ProximitybasedpositioningifBLEtag’spowerislow

WLANaccesspoints

BLEtags

Location

Location

FINGERPRINTINGLOCALIZATION

Deterministicfingerprinting

Probabilisticfingerprinting

LocationKNNNN

WKNN

Coveragearea

Path-loss

Machinelearning

SVM

Decisiontrees

Fig. 2. Approaches for RSS based positioning.

measuring its distances to several reference points. First, theRSS measurements are converted to distance from a mobilenode to APs, then the location of a node is computed usingthe geometry of circles. However, in the case of WLAN basedpositioning the trilateration method does not give accurateresults because of the errors arising when converting RSSmeasurements to distance. This problem is especially severein indoor environments where WLAN signals can be reflectedby walls and attenuated by people’s bodies. In addition, RSSmeasurements are also affected by the smartphone orientation.The conversion of signal strength to distance is based on apath-loss model such as Eq. 6 which will be discussed insubsection II-D2.

An alternative to explicit conversion of signal strengthmeasurement into distance is the comparison of the RSSmeasurements with a radio map. This pattern recognitionapproach to localization is called location fingerprinting. Thereare two stages for location fingerprinting (Fig. 4): offline stageand online stage. In the offline stage, the site is surveyedand the RSS values ("fingerprints") are measured at different

!

B!A!

C!

RA!

RC!

RB!

Fig. 3. Trilateration using signal strength (range) measured from three AP.

3

POSITIONINGPHASE

TRAININGPHASE

AP1AP2...APN

AP1AP2...APM

RSSmeasurementsatcalibrationpoints

RSSmeasurementsatunknownlocations

RSSfingerprints(radiomap)

Fingerprintinglocalization

APsselection

Userlocation

Fig. 4. Location fingerprinting approach based on RSS measurements of thesignals from WLAN access points and BLE beacons.

calibration points throughout the area of interest and recordedin a database. The online stage is the user positioning, whenRSS measurements on the mobile device are compared withrecorded fingerprints from the database. The set of "heard"APs at some particular location during the online stage can bedifferent from a set of APs stored in the radio map becausesome APs can be removed and new APs added or transmittedpower settings can change. Therefore, an AP selection pro-cess is included in online positioning. The location that bestcorresponds to the observed signal strengths is recognized asthe user location. A challenge with fingerprinting is that thepreviously recorded RSS values may be quite different fromthe current values because of changes in the environment andin settings in APs. Removal or addition of new APs in the areaof interest would require modifications in the database. TheRSS measurements can also be affected by the type and modelof mobile device. Despite these limitations, the fingerprintingtechnique can achieve reasonable accuracy and has become amainstream approach to WLAN based positioning.

Fingerprinting algorithms can be deterministic or probabilis-tic. In deterministic matching, one or several survey locationsare found that best match the observed RSS values withrecorded fingerprints. In probabilistic matching, signal strengthvalues are represented as a probability distribution, and thealgorithm calculates the probability of a user’s location basedon the online measurements and RSS database.

B. Radio mapThe radio map is created as a part of the offline learning

phase and it contains the measured RSS values at known loca-tions called calibration points. The advantage of this approachis that the characteristics of the signal propagation in indoorenvironment can be captured avoiding difficulties of modelingcomplex signal propagation. However, the data collectionrequired for radio map generation is a time consuming andlabor intensive process. The radio map generation starts bydividing the surveyed area into cells based on a floor plan. Foreach cell, the RSS transmitted by different APs is measuredover a certain time period at a calibration point and stored inthe radio map database. The i-th element in the radio map canbe represented as [2], [3]:

Mi = (Bi, {aij |j ∈ Ni}, θi), i = 1, . . . ,M, (1)

II. RADIO MAP

Radio map F is the set of all fingerprints, that is

F =!

i∈IF

Fi,

where IF is the set of all possible indices and Fi is the ithfingerprint. The ith fingerprint has the form

Fi = (Ai, ai, Pai) , (2)

where Ai ⊂ Rd is the ith cell, and the approximate coordinateof the ith calibration point is denoted by pi. The jth compo-nent of the vector ai is the mean value of the RSSI valuesmeasured from the access point APj and the jth diagonalelement of the diagonal matrix Pai is the variance of the RSSIvalues measured from the access point APj . It is possiblethat a fingerprint contains more information that might beuseful than what is presented here, see examples [17, 18, 19].However, the above mentioned information is sufficient forFKF. An illustrative example of a possible radio map for asingle access point is presented in Figure 1. The red polygonsrepresent the cells and the stem plot describes the means (a)of the RSSI readings recorded in each calibration point (p).In this particular example, certain cells are empty because noRSSI readings were obtained from the access point in thesecells.

APSignal strength

Figure 1. Example of a radio map

We assume that the measurements collected in the calibra-tion point pi or at least inside the current cell Ai represent theRSSI spectrum inside the cell. This means that in every pointof the cell Ai we can use quantities ai and Pai as the estimatesof the mean of RSSI measurement and the covariance ofRSSI measurement, respectively. This assumption holds ifwe have small enough cells. We also assume that cells Ai

are disjoint, that is Ai ∩ Aj = ∅ if i = j. Furthermore,we assume that the cells cover all possible locations, that isP (∪i∈IAAi) = 1 for every time instant. This is quite a strong

assumption and is not always true. One possibility for thecases where we have an incomplete radio map is to interpolatethe missing fingerprints using the existing fingerprints orpredict the fingerprints using radio wave propagation models.However, due to the complexity of indoor environments, fillingthe gaps in the radio map might turn out very challenging.More information about these approaches are in publications[8] and [20].

III. BEST LINEAR UNBIASED ESTIMATOR

In this section, we present the well-known Best LinearUnbiased Estimator (BLUE) [5, 16], which is the heart ofthe new filter (Sec. IV). In this section we have omitted thesubscripts since we are examining a static estimation problem.However, in the next section we will use the estimator recur-sively and thus denote the time index with a subscript. Let theexpectation and the covariance of the joint distribution of thestate vector x ∈ Rnx and the measurement vector y ∈ Rny

be

E

"#xy

$%=

#xy

$and (3a)

V

"#xy

$%=

#Pxx Pxy

Pyx Pyy

$, (3b)

respectively. All linear unbiased estimators of the state x havethe form

x = x + K(y − y),

where the gain matrix K ∈ Rnx×ny is arbitrary. The covari-ance of the estimator error is

VK△= E

&(x − x) (x − x)

T'

= Pxx − KPyx − PxyKT + KPyyK

T

= Pxx − PxyP−1yy Pyx

+ (KPyy − Pxy) P−1yy (KPyy − Pxy)

T .

(4)

The Best Linear Unbiased Estimator is such that it minimizesthe estimator error covariance (4). It is easy to see (using (4))that the minimun is obtained by choosing the gain matrix as

KKF = PxyP−1yy .

This is the same gain matrix as the gain matrix of KF. Herethe minimization means that the matrix

VK − VKKF = (KPyy − Pxy) P−1yy (KPyy − Pxy)

T ≥ 0

is positive semidefinite for all K. So, BLUE and the corre-sponding covariance of the error are

x = x + PxyP−1yy (y − y) (5a)

VKKF = Pxx − PxyP−1yy Pyx (5b)

Furthermore, it can be shown that if the joint distribution ofthe state x and the measurement y is Gaussian, the conditionaldistribution of the state is Gaussian, and BLUE (5) computesthe parameters of that distribution [21, Theorem 3.2.4] [5].

Fig. 5. Example of a radio map. (Reprinted from [4])

where Bi is the i-th cell centered at i-th calibration point,array aij consists of the RSS measurements from APj , Ni isthe set of all transmitting APs that can be heard at the ithcalibration point, and the parameter vector θi contains anyadditional information that can be useful during positioningphase, for example, antenna orientation, time of day, etc.

The RSS values may be affected by the orientation andlocation of the smartphone with respect to the user’s bodyduring the measurements, device heterogeneity and also by thedensity of people within the building at this time, since thehuman body strongly attenuates signal at 2.4 GHz frequencies.The radio map accuracy also degrades over time becauseof the WLAN APs transmitting power variation and otherenvironmental changes such as position of furniture and walls.Therefore, new fingerprints have to be collected periodicallyto keep the radio map up to date.

C. Deterministic fingerprinting algorithms

Most of the deterministic fingerprinting approaches computeposition based on minimization of the Euclidean distance inthe RSS measurement space considering all "heard" APs [5].The user location estimate is chosen to be the calibration pointxj that is closest (Eq. 2), that is

x = argminxj

(n∑

i=1

(ri − ρi(xj))2)

(2)

where ri is the measured RSS from the i-th AP duringthe online positioning phase, ρi(xj) is the i-th AP’s RSSfingerprint measured at the j-th calibration point during thesite survey phase.

Examples of deterministic location fingerprinting algorithmsinclude the nearest neighbour algorithm and its generalizationsKNN and WKNN [3], [6], [7]. These methods determine thelocation of the mobile node by comparing the recorded RSSvalues from the database with the RSS values measured bya mobile node. k calibration points with the most similarRSS value sets are selected. In this approach, k is a designparameter that can be selected to achieve better accuracy.Usually, when the radio map density is high k should be small.For typical cases the optimal k value is about 3-4.

The estimated user location is computed giving equal im-portance to all k fingerprints. A generalization is to assign

4

different importances to fingerprints, which is known asweighted KNN. The common choice for weighting is to givelarger weights to the fingerprints that are closer in RSS space.The nearest neighbour algorithms are non-parametric methods.Their advantages are simplicity, robustness and reasonablygood accuracy.

D. Probabilistic fingerprinting algorithms

Many approaches apply probabilistic fingerprinting algo-rithms [5], [8], [9] by estimating the probability distribution ofthe measured RSS values given the value of the user’s locationp(x|r1, . . . , rn). The most likely user’s location maximises theposterior probability. Therefore, the problem can be formu-lated as

x = argmaxx

(p(x|r1, . . . , rn)) (3)

This so-called posterior distribution of the location distributionis usually computed using Bayes rule:

p(x|r1, . . . , rn) =p(r1, . . . , rn|x)p(x)

p(r1, . . . , rn)(4)

The prior distribution p(x) encompasses additional informa-tion about personal user profile, for example, tendency towardssome particular location [8]. In most publications it is assumedthat there is no a priori information about the user position andall locations can be accessed with equal probability, i.e. theprior is uniform. The denominator also does not depend on thelocation and can be treated as a normalizing constant. In thiscase the problem is equivalent to the Maximum Likelihoodestimation

x = argmaxx

(p(r1, . . . , rn|x)) (5)

The solution of this problem requires knowledge of thelikelihood function that describes the distribution of the finger-print data at all possible locations. The likelihood function canbe approximated by histograms or some parametric distribu-tions, for example Gaussian distribution. This leads to non-parametric or parametric probabilistic methods. Parametricor model-based algorithms use models with small numbersof parameters to describe the fingerprint data. Compared tonon-parametric methods they have one major advantage –reduced data transmission. Two widely used examples ofprobabilistic parametric location fingerprinting algorithms arecoverage area models and path-loss models [3], [7], [10]. Theformer approach uses elliptical probability distributions formodeling the area in which an AP’s signal can be received.The latter is based on the RSS power loss model calibratedfrom fingerprinting data.

1) Coverage area models: A coverage area (CA) estimateis a probabilistic model of the set of locations at which signalsfrom this specific AP was received in the learning phase. Itis not a model of the actual geographical CA of WLAN AP(unless the survey points were uniformly distributed), and itdoes not model the AP location. If fingerprint location densitymatches user location density (which is reasonable assumptionfor fingerprints collected by crowdsourcing), the CA estimatemodels the location probabilities of the users within receptionrange of the WLAN AP, although the actual reception area

!

http

://m

ath.

tut.fi

/pos

grou

p/

TA M P E R E U N I V E R S I T Y O F T E C H N O L O G YI n s t i t u t e o f M a t h e m a t i c s

POSITIONING WITH BAYESIAN COVERAGE AREAESTIMATES AND LOCATION FINGERPRINTS

Laura Koski, Robert Piche, Simo Ali-LoyttyTampere University of Technology, Finland

ABSTRACT

We present a method to estimatethe coverage area of a wirelesscommunication node (CN), e.g.cellular base station or WLANaccess point, and a methodto use a database of suchcoverage areas for personalpositioning. The coverageareas are modeled as ellipsoids,and their location and shapeparameters are computed fromreception samples (”fingerprints”)using Bayesian estimation. Theposition is computed as aweighted average of the ellipsoidcenters, with weights determinedby the ellipsoid shape parameters.The method is compared toexisting methods using simulateddata.

COVERAGE AREA ESTIMATION

Coveragearea

Fingerprints

The coverage area is estimatedfrom reception samples in thecoverage area. Measurementmodel is

Y = 1nµT + ϵ,

where ϵ(i)|µ, Σ ∼ Np(0p, Σ). Theparameters in interest are µ, thelocation parameter, and Σ, whichdescribes the shape and the sizeof coverage area. µ and Σ areassumed to be random variablesand we compute the posteriordistribution of µ and Σ. Thiscoverage area model models thedistribution of users rather thanthe real coverage of the CN.

Prior

Bayesian prior models informationabout typical coverage areas.

Likelihood

Likelihood describes the knowledgeabout the coverage area giventhe data. Likelihood does not givea good estimate of coverage areawhen only a few samples havearrived. Likelihood can be usedto update the prior information.

Posterior

Posterior distribution combinesthe prior information and informationgiven by the data. Posteriordescribes the distribution of yourposition p, when you hear CN c,p|c ∼ Np(µ, Σ).

POSITIONING

In positioning phase the coveragearea estimates are used as

measurements. The ideais to find the most probablelocation of mobile terminal(MT) given the heard CNs.Given the CNs c1, c2, . . . , cn withknown coverage area estimates((µ1, Σ1), (µ2, Σ2), . . . , (µn, Σn)),the location p of MT is distributedas p ∼ Np(p, C), wherep =

!"ni=1 Σ−1

i

#−1 !"ni=1 Σ−1

i µi

#and

C =!"n

i=1 Σ−1i

#−1.

c2

c1

c3

Distribution ofposition estimate

EVALUATION

Coverage area –based positioningis tested using simulated networkconsisting of cellular networkand WLAN. Coverage area –based positioning is compared toexisting method, which calculatesthe position estimate to be themean of coordinates of CNs.

Errorsmean 68 % 95% RMSE[m] [m] [m] [m]

Coverage135 122 451 261

areaMean of

878 700 2965 1269CNs

100 m

Mean of CNsPosition estimate usingcoverage area models

True position

!

http

://m

ath.

tut.fi

/pos

grou

p/




ABSTRACT



Coveragearea

Fingerprints


Y = 1nµT + ϵ,


Prior


Likelihood


Posterior


POSITIONING



!"ni=1 Σ−1

i

#−1 !"ni=1 Σ−1

i µi

#and

C =!"n

i=1 Σ−1i

#−1.

c2

c1

c3


EVALUATION



Coverage135 122 451 261

areaMean of

878 700 2965 1269CNs

100 m


True position

(a) (b)

Fig. 6. Estimation of a coverage area from fingerprints. (Reprinted from [11])

can be much larger. The CA method considered in [7], [10],[12] ignores the specific RSS values corresponding to IDs ofobserved APs. Therefore, it is more robust to changes in theradio environment than fingerprinting methods that use thesevalues. However, CA methods have lower accuracy comparedwith nonparametric fingerprinting methods such as KNN andWKNN that record corresponding RSS values, and use themduring positioning.

Fig. 6a conceptually shows the actual CA from WLANaccess point, and the fingerprints collected during trainingphase. The CA is approximated by an ellipsoidal probabilitydensity shown in Fig. 6b. The distribution’s location and shapeparameters are computed from fingerprints using Bayesianestimation, which combines the prior information and infor-mation given by the data [12]. Given the coverage areas of allAPs that can be "heard", the user position can be computedas a weighted average of the ellipsoid centers, with weightsdetermined by the ellipsoid shape parameters, as shown inFig. 7.

2) Path-loss models: PL models describe the RSS as afunction of distance from the AP. In the simplest modelsthe PL depends only on the transmission power and distancefrom AP. A standard assumption is that the signal attenuateslogarithmically in terms of the distance between the AP andthe user. A PL model typically contains several parametersthat have to be determined before the positioning phase. Acommonly used PL model [13] is

RSS = A− 10n log10 ‖x−m‖+ w (6)

where A is apparent transmission power, n is a parameterdescribing attenuation properties of the environment, m is APlocation, x is a mobile device’s location, and w is a zero-mean Gaussian random variable used for modeling the shadowfading. Positioning based on PL model has two phases: anoffline learning phase and online positioning phase. In thelearning phase RSS values are collected at known locations.Then unknown parameters A, n, m are inferred for each APfrom the posterior probability distribution [14]

p(A,n,m|RSS1:n) ∝ p(RSS1:n|A,n,m)p(A,n,m).(7)

Nurminen et al. [14] estimated AP position and PL modelparameters simultaneously using the Iterative Reweighed LeastSquares method and Bayesian approach, which permits updat-ing AP position estimate and PL model parameters as newfingerprint data becomes available. During positioning phase

5

http

://m

ath.

tut.fi

/pos

grou

p/




ABSTRACT



Coveragearea

Fingerprints


Y = 1nµT + ϵ,


Prior


Likelihood


Posterior


POSITIONING



!"ni=1 Σ−1

i

#−1 !"ni=1 Σ−1

i µi

#and

C =!"n

i=1 Σ−1i

#−1.

c2

c1

c3


EVALUATION



Coverage135 122 451 261

areaMean of

878 700 2965 1269CNs

100 m


True position

Fig. 7. User position estimation using coverage areas. (Reprinted from [11])

the user position is inferred from the RSS that can be detectedfrom several APs according to the following equations

RSSi|x, Aj , nj ,mj ∼ N(Aj − 10nj log10 ‖x−mj‖ ,Σ)p(x, A1:k, n1:k,m1:k|RSS1:k)∝ p(RSS1:k|x, A1:k, n1:k,m1:k)p(x, A, n,m)p(x|RSS1:k) =∫p(x, A1:k, n1:k,m1:k|RSS1:k)dA1:kdn1:kdm1:k

(8)where Aj is apparent transmission power from j-th AP, nj isa parameter describing attenuation properties of the environ-ment, mj is j-th AP location, x is a mobile device’s location,and Σ is a covariance of RSS measurement noise. From RSSmeasurements and PL models the distances between a set ofreference nodes and the target node are estimated, which thenenables estimation of the user’s position. However, the positionestimate is sensitive to signal noise and PL model parameteruncertainties, which are large because the distance-powergradient is relatively small. Consequently, these estimates aregenerally less accurate than radio signal based estimates thatare derived using angle of arrival or time delay measurements.

E. Application of machine learning methods and compressivesensing in fingerprint positioning

Support vector machines (SVM) is a technique for dataclassification, statistical analysis and machine learning thatwas recently applied for RSS fingerprinting positioning [15],[16]. The SVM classifies data by finding the best hyperplanethat separates all data points of one class from those of theother. The best hyperplane for an SVM means the one withthe largest margin between the two classes. Margin means themaximal width of the parallel to the hyperplane that has nointerior data points. Classification of the fingerprint databasecan obtain optimal nearest neighbor points. Wu et al. [15]mentioned that when SVM is applied to RSS fingerprintinglocalization the inference model has to be built in advance.Therefore, SVM does not solve the biggest disadvantage oftraditional RSS fingerprinting related to adaptivity to changesin the environment. Feng et al. [16] proposed to use theSVM and bilinear median interpolation method to reducethe calibration effort on creating RSS fingerprint map whileretaining the required positioning accuracy.

Sanchez et al. [17] proposed an indoor localization systembased on multiple weighted decision trees. In addition to theRSS from different APs, the training dataset includes thedevice orientation because the RSS is significantly influencedby the different device orientations. This approach is com-putationally efficient and can be implemented on a resource-constrained mobile devices. This localization algorithm isabout 100 times faster than well known Horus [9] RSSfingerprinting positioning algorithms.

Feng et al. [18] applied the compressive sensing approachfor the radio map reconstruction using only a subset of RSSfingerprints, thus significantly reducing the number of mea-surements during fingerprinting database update. Compressivesensing is a signal processing algorithm for sparse signalsrecovery from a small number of noisy measurements bysolving an L1-minimization problem. Their location estimatorconsists of a coarse localizer, where the RSS is comparedto a number of clusters to detect in which cluster the nodeis located, followed by a fine localization step, using thetheory of compressive sensing, to further refine the locationestimation.

F. Difficulties with WLAN based positioning system deploy-ment

The accuracy of WLAN RSS-based positing systems can beseriously affected by attenuation of WLAN signals by people,heterogeneous mobile devices and their pose. The effective-ness of WLAN based fingerprinting system deployment ischaracterized mainly by the required time and efforts for sitesurvey and maintenance of fingerprint database to keep it up-to-date.

1) Devices heterogeneity and orientation: RSS measure-ments made by different commodity smartphones can varysignificantly even under the same wireless conditions sincethese devices are usually equipped with different WLANchipsets and antennas. Viel and Asplund [19] observed thedifference in RSS of up to 12 dBm from same signal measuredby different smartphones. According to He and Chan [20] themajor factors that affect the RSS measurement on smartphonesare WLAN chipset sensitivity, antenna installation position andsmartphone’s operating system. Therefore, a WLAN basedpositioning system that relies on RSS fingerprints generallydoes not perform well across heterogeneous devices.

Hossain et al. [21] presented new approach for RSS indi-cator calibration on heterogeneous devices that is robust tomobile device hardware heterogeneity. The approach is basedon a new location fingerprint, namely, signal strength differ-ence (SSD) that is free from hardware-specific parameters,such as antenna gain and transmitter/receiver power gain. Thealgorithm selects one AP measurement which will be deductedfrom all other RSS measurements. Therefore, the constantfactor of antenna gain is deducted. However, this approachsuffers from RSS noise fluctuations. The SSD fingerprintingis also applicable to BLE based positioning.

There are also approaches for offline calibration of het-erogeneous smartphones. Laoudias et al. [22] noticed thatRSS measurements on different devices are correlated and can

6

be approximated by a linear model. They proposed to builda linear regression for RSS measurements on two differentdevices. The disadvantage of this method is that it requiresoffline training before an accurate mapping relationship canbe learned.

Chapre et al. [23] measured RSS on a stationary laptopfrom 13 APs in a typical office environment. The data wascollected for 4 directions such as east, west, north and southduring 24 hours at a fixed location. The results show thatthe RSS variation for the different directions is on average 3dBm. Della Rosa et al. [24] conducted similar experiments onsmartphones and also reported that RSS measurements dependon the specific smartphone’s orientation with respect to the AP.The RSS variation for different directions depends also on thetype of smartphone. It was 3.6 dBm for Nokia N900 and 7.6dBm for iPhone 4.

2) Attenuation of WLAN signals by people: Della Rosa etal. [25] studied the body-loss effect on the RSS measurementson smartphones. The experimental results show that the signalstrength at a given location varies by up to 5 dBm dependingon the direction that the user is facing because the user’s bodycreates a systematic source of error in RSS measurements.Correcting this source of error during positioning phase can bedifficult. Hand grip is another source of error in the RSS mea-surements on smartphones because of close proximity of handto the antenna. In [25] the authors made RSS measurements ona smartphone in the same environment for two different cases:(a) the hand is not covering the WLAN antenna and (b) theantenna is fully covered. The test results show that the RSSerror because of hand grip can reach 12 dBm that correspondsto approximately 9 m of range error at the distance of 3 mfrom the AP.

3) AP selection: Appropriate AP selection can reduce thecomputational burden and in some cases improve the WLANbased positioning accuracy. Youssef et al. [26] proposed toreduce computational cost by choosing only a subset of all"heard" APs with the strongest signal. In this case the positionis computed using only information from the selected APs andother APs are discarded.

Fang and Lin [27] developed another approach for reductionof fingerprint dimension transforming RSS to a new set ofvariables that is comprised of principal components (PC). Inthis way the information for different APs is reorganized andsqueezed into lower dimension by a linear transformation. Thework also presents an algorithm for determining the number ofPCs required to obtain good performance and make sure thatthe discarded information is in fact the redundant noise. Heand Chan [20] assumed that a major concern with fingerprintdimension reduction proposed in [27] is that the selected APsand their weights derived from the offline training samplesmay be not optimal during online positioning because theenvironment dynamically changes. The importance of APs inonline localization can be different if the set training samplesis too small or outdated.

G. Positioning accuracy of WLAN based approachThe positioning accuracy of WLAN based approaches de-

pends mainly on the following factors: density of the WLAN

APs, resolution of the radio map, degradation of radio mapaccuracy with time, and positioning algorithm. Therefore, thepositioning performance of the same algorithms can be verydifferent at different sites, or even in the same building if theradio map has not been updated for long time and collectedfingerprint data does not represent well actual signals. Mülleret al. [7] evaluated major algorithms for WLAN based posi-tioning by analyzing their positioning accuracy during fieldtests in a university campus that represents a typical officebuilding densely covered by WLAN APs. The goal of theseexperiments was to compare the performance of different fin-gerprinting algorithms: WKNN, coverage area (CA), path-loss(PL), and generalized Gaussian mixture (GGM) approximationof the PL model. The major effects of environmental changessuch as density of APs and radio map accuracy degradationwere also evaluated.

An example of estimated user position for three parametricfingerprinting algorithms and an up-to-date radio map is shownin Fig. 8. The actual user location is shown by a blackline and the recorded fingerprints are depicted as grey dots.The position accuracy was slightly better for PL and GGMalgorithms: 50% of time the error was below 8 m, compareto CA algorithm for which the error was about 10 m. How-ever, CA algorithm requires 30-50 times less computationaleffort compare to PL and GGM. Non-parametric WKNN alsoshowed good performance with position error below 6 m50% of time. When 90% of access points were dropped theposition error of CA algorithm slightly increased to about12 m (50% of time). The GGM algorithm’s position errorwas also about 10 m, and PL algorithm’s position error wasabout 15 m. However, the performance of WKNN algorithmdeteriorated significantly exhibiting the largest position errorsamong all of these algorithms. Finally, the performance ofthese algorithms was also compared when the fingerprints datawas not up-to-date: data for generating the radio maps anddata for positioning were collected with a time gap of severalmonths to evaluate the performance degradation over time. Allparametric methods preserved their positioning performances,while the WKNN algorithm’s performance degraded signifi-cantly [7].

Fig. 8. Field test results for three commonly used WLAN based positioningalgorithms. (Reprinted from [7])

7

H. Future trends in WLAN based positioning

Unlike the RSS values that essentially represent the totalreceived power, the channel state information (CSI) containsinformation about the channel between sender and receiverat the level of individual data subcarriers for each pair oftransmit and receive antennas and reports the amplitude andthe phase of each of the subcarriers [28]–[30]. The CSIdescribes the combined effect of scattering, fading, and powerdecay with distance. It achieves higher robustness than tradi-tional RSS fingerprinting and it can potentially replace RSSin WLAN based positioning systems. However, the WLANcards on modern commodity smartphones do not supportthe CSI data collection. Current prototype systems for CSIdata collection are laptop based and use the Intel WiFi Link5300 network interface card, which is the only commercialtool for CSI data recording based on the 802.11 standard.The example of WLAN fingerprinting positioning systemsare given in [31], [32]. Wang et al. [31] proposed a deeplearning based fingerprinting scheme that uses CSI informationfor all the subcarriers from three antennas. Another versionof fingerprinting positioning approach is based on calibratedphase information of CSI [32].

I. BLE based positioning

The emergence of Bluetooth Low Energy (BLE) beaconsopens up a new generation of indoor positioning systemsthat can be more accurate, reliable and efficient. BLE can beused in new applications for a multitude of new markets suchas IoT, connected home, health, omnichannel retail, ambientintelligence, augmented reality, and mobile advertising. ABIResearch’s report, "BLE Tags: The Location of Things" statesthat total BLE beacon shipments could exceed 400 millionunits in 2020. The Bluetooth Special Interest Group predictsthat by 2018 more than 90 percent of Bluetooth-enabledsmartphones will support the BLE standard, so this technologywill become ubiquitous.

BLE devices operate in the same 2.4 GHz band as WLANand Classic Bluetooth 2.1. However, BLE has some specialcharacteristics such as high scan rate, short handshake proce-dure and very low power consumption that make it differentfrom these technologies. Since BLE beacons are powered bytheir own batteries they can be placed everywhere, unlikeWLAN AP, which are typically placed near power outletsin such a way as to optimize signal coverage with minimalinfrastructure deployment. So, another advantage of BLE isin the freedom to place beacons to provide good signalgeometry [33].

1) BLE beacons: At the moment, there are several typesavailable, each with their own standards and advantages. Themost popular BLE beacon ecosystems are Apple’s iBeacon,Google’s URIBeacon and Eddystone, and Radius Networks’AltBeacon. iBeacon reliably works with Apple’s iOS andcompatible with Google’s Android. Other beacons have opensource protocol and are compatible with different mobileoperating platforms. A BLE beacon is self-contained and verysmall, roughly the size of a coin and it is in sleep modemost of the time and only wakes up when a connection is

initiated [34]. Power consumption is kept low because theactual connection times are only a few milliseconds. Themaximum power consumption is 15 mA, and the averagepower consumption is only about 1µA. A BLE beacon isusually a 1-way transmitter that transmits a 16-24 digit stringof characters, which can identify the individual beacon.

2) Positioning with BLE beacons: Indoor positioning basedon BLE can use the same methods as WLAN based posi-tioning, namely, fingerprinting. One difference is that BLEis available in class 1 and class 2 versions, where class 1has a data transmission range of up to 100 meters in openspaces while class 2 range is about 10 m. Radio maps forBLE can include one additional parameter describing the classof Bluetooth. Zhao et al. [35] compared BLE and WLANbased positioning and came to conclusion that BLE is superiorfor indoor localization for the following reasons: (a) channelhopping mechanism, (b) lower transmission power, and (c)much higher scan rate than WIFi. Frequent channel hoppingcan average out interference in a given channel or completelyremove interference when BLE radios hop to the next channel.This makes the RSS less noisy. The lower transmission powerof BLE can reduce the multipath effect in some scenarios. Thehigh scan rate makes it possible to average out the occasionaloutliers caused by interference or multipath effect and improvethe tracking accuracy.

Another difference is that BLE can provide very accurate(submeter-level) positioning in proximity mode when thetransmitting power is set to low levels. The proximity isusually detected applying thresholds on RSS measured on asmartphone. To reduce fluctuations in the measured RSS thedata can be smoothed, for example, by applying exponentiallyweighted moving average [36]. Apple and Qualcomm havereleased frameworks that exploit the proximity mode of BLEto provide "microlocation" and detect the presence/absenceof a device within a certain radio range. Based on this,notifications and special offers can be sent to the user’ssmartphone.

In close proximity (less than one meter) to a beacon thedistance estimation based on RSS is very accurate because thesignal power decreases as the lognormal attenuation model,which relates the RSS to distance d with the followingfunction [35], [36]:

RSS(d) = RSS(d0) + 10n log(d/d0) +Xσ (9)

where RSS(d0) is the value of RSS (dBm) at the referencedistance d0, n is the path loss exponent, and Xσ is thezero-mean Gaussian noise with variance σ2. The path lossexponent has a theoretical value of 2 for unobstructed freespace propagation, corresponding to a loss of -20 dB/decade.However, in real environments path loss typically is muchhigher due to shadowing and multipath interference. Thereforethe ranging performance deteriorates with range, and at 10 mthe ranging error would be around 5 m [33].

3) Accuracy of BLE based positioning: Zhao et al. [35]conducted extensive experiments in indoor environments andcompared the positioning accuracy of BLE and WLAN in al-most identical environment and conditions. The results showedthat BLE based positioning is more accurate than WLAN

8

by around 27 percent. The authors also demonstrated thatBLE propagation model can describe RSS as function ofrange better than WLAN. This can be explained by the lowertransmitter power of BLE and, as a consequence, smaller rangeand less multipath. This is the major reason why BLE is bettersuited for positioning than WLAN and potentially can be moreaccurate when used in indoor navigation.

Lohan et al. [37] also compared positioning accuracy ofWLAN and BLE based positioning. In their setup the numberof WLAN routers was almost three times larger than thenumber of BLE beacons. So, the environment and conditionswere not identical. The results showed that the positioningaccuracy was almost the same for WLAN and BLE. Theauthors made a general conclusion that the WLAN and BLEsignal propagation in multi-floor buildings is very similar, andthat a PL model with a floor-loss factor can be successfullyemployed in order to achieve good floor detection probabilitiesand low position estimation errors.

J. Summary of WLAN and BLE based positioningCurrently, WLAN and/or BLE based positioning is a back-

bone for indoor navigation with a smartphone. The advantageof this technology is in the wide dissemination of WLANinfrastructure and the growing number of BLE beacons andalso the presence of integrated WLAN and BLE radio inalmost every smartphone. The main drawback is radio mapcreation and maintenance. One solution to this problem isto use crowdsourcing. However, how to implement this isstill an open problem. Other disadvantages of WLAN basedpositioning are the low scan rate of WLAN RSS, usually0.2-0.3 Hz and large errors in RSS measurements due todevice heterogeneity, smartphone’s orientation, and attenuationof signals by people.

In typical office buildings or shopping malls WLAN basedpositioning provides an accuracy of about 10 meters, whichis not always enough to meet the requirements for LBSservices of a room level place recognition and correct flooridentification. The BLE positioning can be potentially veryaccurate with up to one meter position accuracy in proximitymode. However, the BLE beacons are not so ubiquitous asWLAN APs. The WLAN based positioning can be improvedif it is aided by other technologies such as map-matching andself-contained sensors.

III. MAGNETIC FIELD FINGERPRINTING

Magnetic field fingerprinting uses a map of magnetic fielddistribution inside buildings for indoor positioning. Distur-bances of the Earth’s magnetic field indoors are caused mainlyby the metal structure of buildings. The positioning approachconsists of two phases [38], [39]: offline mapping of themagnetic field measurements at known locations and onlinepositioning by matching the measured magnetic field withthe fingerprints from the database. Unlike WLAN or BLEbased positioning, magnetic fingerprinting provides only localpositioning because of its spatial ambiguity.

This approach has one important advantage that makes itattractive for indoor positioning: the magnetic field is every-where, it is relatively stable [40]–[42], and consequently no

pre-installed infrastructure is required. Since the output ratefor a typical smartphone’s magnetometer is about 10 Hz thegeneration of magnetic field map and positioning can be con-tinuous [43]. However, there are also serious disadvantages:(a) a single fingerprint consists of only a small number ofparameters, at most three, but usually two or even one; (b) themagnetic field gradient sometimes can be very steep [40]; (c)intermittent magnetic interferences can be significant.

Measurements of the magnetic field vector come from thesmartphone’s 3D magnetometer in the device coordinate sys-tem. Since the smartphone pose can change freely the knowl-edge of the transformation between the device and globalframes is required. Accurate computation of smartphone’s 3Dpose is a difficult task because the heading measurementsare unreliable when magnetic field disturbances are large. Inmany real life scenarios the direction of gravity (measuredby accelerometers) is the only pose component that can becomputed, thus reducing the number of fingerprint parametersto two [44]. If even the direction of the gravity cannot becomputed, then only the magnitude of the magnetic fieldintensity ‖m‖ can be used as a single fingerprint parametersince magnitude ‖m‖ is independent of magnetic sensorpose [38]. The small dimension of fingerprints makes themagnetic field fingerprinting unreliable and spatially ambigu-ous. In addition to this it also sets stringent requirements forfingerprint database creation. The large gradients of magneticfield are advantageous for positioning making possible sub-meter position accuracy. However, it also requires a very highresolution magnetic field map, thus making efficient magneticfield mapping one of the biggest challenges for this approach.

Other challenges of using the magnetic field for posi-tioning with smartphones include diversity of devices andusage scenarios, and significant variability of magnetic fieldwith altitude. Shu et al. [42] demonstrated that for the samemagnetic field, different devices show different readings ofthe magnitude of the collected magnetic signals along ex-actly the same path. Also for the same device, the datacollected at different device’s altitudes is different. Therefore,to achieve good accuracy the magnetic field map should bethree-dimensional. Also during positioning phase the altitudehas to be accurately estimated which is impossible in the caseof pedestrian navigation with smartphones and usually onlycorrect floor can be identified.

A commonly used approach for incorporating measuredmagnetic field intensity into a positioning system is to usea statistical model. If magnetic field measurement errors areGaussian the likelihood function is given by a multivariatenormal distribution with mean h(x) and covariance R [38],[45]

p(z|x) =1√

(2π)N |R|exp

(−1

2(z− h(x))TR−1(z− h(x))

)

(10)where z is the magnetic field intensity measurement, x is thelocation, and the function h(x) represents magnetic field mapor fingerprints generated during training phase. If the magni-tude of the magnetic field intensity is used as an observation,then the multivariate normal distribution in Eq. 10 is replaced

9

by a single variable normal distribution.The framework for position estimation is similar to terrain

navigation with matching of altitude (for aircraft) or depth (forships) [46]. This approach requires data fusion with a position-ing system that produces estimates of the user’s location. In thecase of indoor pedestrian navigation this can be PDR, WLANand/or BLE based positioning or any combination of thesesystems. The dynamic system model including propagationand measurement equations can be defined as [45]

xt+1 = xt + ut + vt; t = 0, 1, 2, . . . (11a)

zt = h(xt) + et; t = 1, 2, 3, . . . (11b)

where ut is the estimated displacement from the previousposition, xt, at time t, vt is the random error associated withthe navigation system that provided xt, h(xt) is the functionthat maps a specific value of xt to the measurement domain,and zt is the actual measurement collected by the sensor,which has a measurement error, et. For the case when themeasurement noise et is Gaussian, the estimation algorithmcan be the maximum likelihood estimator (MLE) with likeli-hood function given by Eq. 10. The MLE can be combinedwith the position estimate from the propagation equation usingBayes rule and the law of total probability. According to theBayesian view, the posterior probability density function (pdf)p(x0...t|z1...t) contains all the statistical information availableabout the state vector sequence x0...t given the informationin the z1...t measurements. If the probabilistic model of thetransitional density is described by a Markov process of thefirst order, then the posterior pdf can be calculated recursivelyas [47]

p(x0...t|z1...t) ∝ p(zt|xt)p(xt|xt−1)p(x0...t−1|z1...t−1).(12)

where p(zt|xt) is the likelihood of xt given the measurementzt (Eq. 10), and p(xt|xt−1) is the mobility model basedon the propagation equation described in Eq. 11a. However,very often the analytical solution for Eq. 12 is not tractable.Therefore, the problem is usually solved numerically by meansof a particle filter, where the posterior pdf is approximated bya particle set [48], or by a point mass filter, where the densitiesare approximated using a grid of point masses [46].

IV. MAP-AIDED NAVIGATION INDOORS

The idea of using building geometry for reduction ofposition and heading errors in autonomous positioning systemshas been extensively exploited in the last several years. Inthe case of pedestrian indoor navigation, building floor plansrepresent constraints that restrict movements. For example,people cannot walk through walls and floor changes canoccur only via staircases or elevators. The goal of map-aidednavigation is to exploit prior information contained in maps orbuilding plans to improve positioning accuracy. The problemis to do in an automated way what our eyes/brains can do sowell. When an indoor map is available in addition to a WLANor BLE based solution it can improve the positioning accuracy.There are currently three approaches to map aided navigationindoors, all of which can be implemented on smartphones:

!!!!!

Smartphone!Sensors!

Motion!models!

Velocity!and!heading!

computation!

Heading!correction!

Extraction!of!topological!features!

Indoor!map!

Coordinates!Location!on!the!map!

Topological!map>matching!

Map>matching!using!wall!constraint!

Link>node!model!of!building!

MAP>MATCHING!ALGORITHMS!

WLAN!(BLE)!based!position!

Fig. 9. Approaches to map-matching indoors.

probabilistic map matching based on particle filtering usingwall constraints, topological map matching based on link-noderepresentation of a building plan, and reduction of headingerror by comparison with building cardinal heading. Thepurpose of these algorithms is to improve positioning andheading by adjusting the estimated path to the building plan.The major components of an indoor map-matching approachare shown in Fig. 9. In some cases some sensors may not bepresent in a smartphone. In this case estimations of velocityand heading that are necessary for particle propagation in map-matching algorithms can be obtained from other sources, forexample, WLAN based positioning.

A. Map-matching indoors using wall constraints

Indoor versions of probabilistic map matching are based onnumerical solution of the Bayesian filtering equation using al-gorithms known as particle filters [49]–[55]. These algorithmsuse a set of weighted random samples (particles) to representthe posterior density of the unknown position in a dynamicstate estimation framework. The particles are distributed overthe digital building plan where walls represent impassableobstacles. A particle that collides with a wall is excluded fromthe Monte Carlo simulation.

The goal of the Bayesian filter is to recursively solve forthe posterior distribution p(x0...k|y1...k), i.e., the conditionaldistribution of the states x0,...,k at time instants 0, 1, . . . , kgiven the sequence of observations y1, . . . , yk. In a case ofpedestrian navigation it is common to assume that the state isdescribed by a discrete-time Markov process. In other wordsthe state at time step tk depends on the previous state at timetk−1 according to the probabilistic model p(xk|xk−1). Thestate of the Markov chain usually consists of the pedestrian’shorizontal position, heading and, optionally, the floor number:xk = [xk yk ψk fk]

T . Here, the k subscript corresponds tothe tk time instant. The evolution of the horizontal positionin time can be described with the aid of a constant velocitymodel or dead-reckoning equations if velocity and heading arecomputed using phone’s sensors. In many cases, an additivezero-mean Gaussian noise in the speed and heading estima-tions provides a good approximation. Then the particles can

10

be simply sampled from the transitional prior described by

p(xk|x(i)k−1) = N (xk; ak,Σk). (13)

In this formula, the i-th particle is a candidate state vector x(i)

with associated weight w(i) ∈ [0, 1], the mean ak for thenormal distribution of position and heading which is calcu-lated based on the motion model. The standard deviation Σkapproximately corresponds to the uncertainty in velocity andheading. However, Σk can be larger to accommodate not onlythe motion model uncertainty, but also the uncertainty inducedby the approximate nature of the particle filter algorithm [56].

In partially observable Markov chains the entire state cannotbe measured. Usually only stochastic projection of the truestate is available. In the case of indoor pedestrian navigationwith WLAN or BLE based positioning the application of themap constraint expresses the fact that people cannot movethrough the building walls. In the proposed map-matchingscheme, the map information is incorporated into the particle’sweight during update step where the measurement likelihoodis defined as

p(yk|x(i)

k

)=

{0 if there is a wall

p(yk|x(i)

k

)otherwise. (14)

here p(yk|x(i)

k

)is the measurement model corresponding to

the user’s location measurement yk computed by the WLANor BLE based positioning system.

This problem can be stated as the estimation of the sequenceof states x0...k = {x0, . . . , xk} in partially observed Markovsequence given a map of the environment, and subject to themotion model p(xk|xk−1) and measurement model p(yk|xk).The prior probability at t0, p (x0) is assumed to be known.The goal is to find the best trajectory in terms of the minimummean square error (MMSE) criterion. The required posteriorprobability distribution can be expressed according to Bayes’law as

p(x0...k|y1...k) =p(yk|x0...k, y1...k−1)p(x0...k|y1...k−1)

p(yk|y1...k−1)

∝ p(yk|xk)p(xk|xk−1)p(x0...k−1|y1...k−1).(15)

In this particular application, a particle filter provides anatural and intuitive way of incorporating building plan in-formation into position estimation, by applying the directconstraint to each particle as was implemented in [49]–[51],[55]. All these approaches use the same method of buildingconstraints implementation. The differences lie in the parti-cle propagation and measurement update computation. Thesealgorithms also use the smartphone’s sensors to compute theparticle movement. Each particle is propagated in a procedureanalogous to Kalman-type filters and can be represented bythe following steps (Fig. 10):

1) Prediction step: The particles are projected to the nexttime step by sampling from a proposal distribution whichcan be chosen to be the transitional model p(xk|xk−1).

2) Application of wall constraint by reducing a particleweight if it crosses a wall.

Algorithm:1. Particle propagation based on PDR

2. Floor plan weighting

3. WLAN weighting

4. Resampling

z }| {

Nurminen, IPIN 2013, slide / 12

The particle filter combines different measurements

6

1.wall

2. 3. 4.

Fig. 10. Application of wall constraints using particle filter. (Reprintedfrom [57])

3) Update step: Applying measurement update based onproximity of particle’s location to the WLAN or BLEcomputed position. The weights are updated accordingto

w(i)k ∝ w

(i)k−1p

(yk|x(i)

k

). (16)

4) Resampling.Since the particles represent a Monte Carlo approximation

of the posterior probability density function expectation, theuser location can be estimated as their weighted mean. Theuser’s path estimated in this way is optimal in terms of theminimum mean square error. Note that the path computed thisway can cross walls because the particles can be in differentparts of the building.

Khider at al. [53] used different particle propagation modelsto increase the robustness of map-matching algorithm, forexample, the stochastic behavioral model, the diffusion move-ment model and map-enhanced combined model. Later theauthors also addressed floor changes in [58] adapting diffu-sion movement models for 3D case. The vertical movementcomputation module is turned on if the person is detected tobe in any of the stairs areas, because in this case the floorchange is possible. The authors also proposed an intermediatevirtual floor between staircases, which they called x1/2 level,to approximate motion between floors inside the staircase.User motion inside the staircase is represented by an extendedMarkov model. The new floor level is set depending on themodel based information. The vertical speed in the staircaseis adjusted depending on the user climbing up or down. Inthis approach the altitude measurements are not used althoughthey could increase the robustness of the floor transitionestimation. Also elevators or ladders for industrial facilitiesare not addressed.

In [51], [55], [58] the map-matching algorithm was alsoextended with 3D movement in multi-storey buildings. Thevertical movement can be detected by the smartphone’s sen-sors. To monitor a floor change Ascher et al. [55] proposed anew representation of the map with three possible objects:

• Rooms: bounding walls, transitions to other objects• Stairs: inclined planes with bounding lines and transitions

to other objects• Ladders/elevators: vertical rectangle with bounding lines

11

Fig. 11. Link-node representation of a building. Links are shown by bluelines.

Transition between floors is only possible through stairs,ladders and elevators. Vertical movement is detected usingbarometer and inertial sensors.

B. Topological map-matching indoors

Similar to the street map-matching the topological algo-rithms for indoor applications use a link-node representationof a building plan [59]. A node is a point defined by its coor-dinates including altitude. The altitude can be given in termsof a floor number. The nodes correspond to the junctions andto the points of interest in the building. A link is a straight lineconnecting two consecutive nodes. The links may correspondto the corridors, staircases, entrances to elevators, passagesbetween buildings, etc. The consecutive links are connectedand represent a network similar to a road network (Fig. 11),where link-node representation of building is depicted byblue lines and dots. The link-node model can be generatedautomatically from a building floor plan using, for example,Dijkstra’s algorithm [60] or Voronoi-Diagram [61].

However, developing a set of map matching functions forpedestrian navigation is a challenge because the trajectory ofpeople is not always similar to the geometry of the mappingdata. Development of algorithms indoors is based on thecomparison of topological elements from the trajectory andthe database. The goal is to identify the correct link and thenlocation of a user on this link. A topological map matchingfor indoor applications relies on the similarity of the trajectorygeometry and the topological features of the link-node graph.The approach described in [59] is based on Bayesian inferencewhere the estimation is computed considering the traveleddistance and azimuth. Spassov [62] used the Fréchet distanceas a degree of similarity between two polylines.

Another approach is to use a particle filter in which theparticles always stay on links [60], [63]. Propagating theparticles on the links imposes an additional constraint onparticle movement and allows the use of a smaller number ofparticles. However, this approach provides an adequate modelof pedestrian movement only in places like corridors wherethe most likely route will be along the corridor. Using a link-node graph it is difficult to model pedestrian movements inlarge open spaces where people can stay and walk anywhere.In practice, better results can be achieved when a link-node

model of building is combined with wall constraints [60]. Inthis implementation corridors and small rooms are representedby graphs whereas wall constraints are applied for large openspaces.

Hilsenbeck et al. [61] proposed to use a grid-based Bayesinfilter in which the continuous state space is replaced bya sparse graph structure that represents a link-node modelof a building and the posterior probability distribution iscomputed only in the selected nodes. Their results demonstrateat least tenfold reduction of computational burden comparedto state-of-the-art particle filter solutions while ensuring robustlocalization. Additional reduction of computational complexitycomes from a coarse, adaptive quantization of the indoorspace, which takes into account typical motion patterns ofpedestrians in buildings. The proposed approach is able toreliably track a pedestrian through an indoor environmentusing only 50 particles to process pedometer measurementsin combination with WiFi signal strength measurements. Inaddition, the algorithm is able to adapt to characteristic gaitpatterns for individual users. The approach also addressednavigation in open spaces by adaptively constructing the graphwith respect to the available space.

The performance of a positioning system based on WLANand building floor plan is shown in a supplementary down-loadable video Clip 1. On this video the true user’s position isshown by a cyan diamond, the WLAN based positioning byblack asterisk, and the position estimated as a result of fusionof WLAN with map is shown by a red circle. The particlesthat are propagated using link-node model are depicted as bluedots. The particles that are propagated using wall constraint aredepicted as red dots. The field tests results indicate that most oftime the map-aided positioning is more accurate than WLANonly positioning. Note also that link-node map-matching betterpreserves the topology: the estimated position follows thetrue path without crossing the walls and going to the wrongplaces. Movement between floors is also implemented usingthe vertical link representing elevators and staircases.

C. Heading correction through the knowledge of the buildingorientation

The algorithms for heading correction with the aid of amap fall into two groups: Algorithms that explicitly generateheading measurements from the knowledge of the buildingorientation; Algorithms that exploit the information containedin building floor plans through map-matching.

Abdulrahim et al. [64], [65] proposed the approach togenerate heading measurements from the basic knowledge ofthe orientation of the building in which the navigation systemis operating. The solution is based on the assumption that mostbuildings are constructed with a rectangular layout where mostrooms and corridors are also rectangular, thus constraining thedirection in which the user can move throughout the buildinginto one of four principal headings. The disadvantage of thesemethods is that a user can actually move in a different directionthan the “cardinal heading” of the building. Therefore themethod may fail in the following cases [64]: (a) a person iswalking in circles or curvilinear lines for long period of time,

12

(b) the building does not conform to the simple geometry, and(c) internal rooms and corridors are not parallel to externalwalls.

Borenstein and Ojeda introduced a method called “Heuris-tic Drift Reduction” (HDR) [66], [67]. HDR makes use ofthe fact that many streets or corridors are at least partiallyrectilinear [66]. At any moment, the HDR method estimatesthe likelihood that the user is walking along a straight line; ifthat likelihood is high, HDR applies a correction to the gyrooutput that would result in a reduction of drift if indeed theuser was walking along a straight line. If the algorithm assessesthat the user is not walking along a straight line, then HDRdoes nothing. The limitation of the HDR method is that whennot moving straight, at best we can expect HDR to notice thatand suspend its operation [67]. During that time, gyroscopedrift accumulates and the integration of the rate of turn resultsin heading errors. Then, when moving straight again, newheading errors are prevented, but those heading errors thatwere accumulated while HDR was suspended remain in thesystem and cannot be eliminated. The effect of the HDRmethod is thus that it reduces heading errors due to driftsubstantially, but cannot totally prevent the unbounded growthof heading errors.

Jiménez et al. [68] analyzed the shortcomings of HDR al-gorithm, which can even degrade the navigation solution whenused in complex buildings where corridors are curvilinearand not aligned to a rectangular layout or where there arelarge open areas. They proposed a method, called improvedHeuristic Drift Elimination (iHDE), that includes a motionanalysis block to detect straight-line paths and an adaptiveon-line confidence estimator for the heading corrections.

In [69] the authors proposed an approach for preventing theunbounded error growth by correcting position and headingerrors of autonomous navigation systems operating indoors.Although this method was applied to vehicle navigation it canbe extended for pedestrian navigation with a smartphone. Thealgorithm comprises three steps: (a) the autonomous sensordata is processed to obtain position, velocity, and attitude,(b) map-matching corrections are applied to the trajectorycalculated by the dead reckoning system, and (c) the mostaccurate estimation of vehicle’s path is computed as optimalfusion of map-matching and dead reckoning solutions.

D. Summary of map-aided positioningTwo different approaches to the fusion of map data with

WLAN based positioning were studied: particle filter with wallconstraints, and particle filter with combined representation ofbuilding plan (link-node and wall constraint). In both casesthe performance is better than performance of only WLANbased positioning. A brief summary of pros and cons foreach map-matching approach is given below. Map-matchingbased on particle filter with wall constraints has the followingproperties:

• Floor plans are readily available• Implementation is relatively simple• The computed user’s path can cross walls• In the case of only WLAN based position estimates, the

propagation of particles may be not accurate

!

Activity!recognition!

Smartphone!sensors!!

Orientation!tracking!

Mode!of!transit!

Step!parameters!

Step!count!Compass!direction!

Gyros!Accelerometers! Compass!

Step!detection!

Step!length!estimation!

Step!direction!

Motion!classification!

Floor!changes!!

Fig. 12. Major tasks that can be performed using smartphone sensors.

Map-matching based on particle filter using link-node repre-sentation of buildings has the following properties:

• Node-link model of building has to be obtained• Implementation is complex, but more stable numerically

and requires small amount of particles• The topology is better followed, however it is not well

suited for large open spaces• Work well even when only WLAN based position esti-

mates are available

V. USING SMARTPHONE SENSORS FOR POSITIONING

Inertial and other self-contained sensors can improve smart-phone based positioning. However, the accuracy of these sen-sors in a typical smartphone is low. Therefore, a conventionalstrap-down inertial navigation approach is not suitable forimplementation because of rapid accumulation of position,velocity and attitude errors. An approach to using smartphonesensors for positioning can be useful only if it can cope withthe changing orientation of a phone without any constrainton usage scenarios which typically include carrying phone ina pocket, bag, hand or arm-band, making a call, texting, orlooking at the display. The major challenge when the phoneis not assumed to be constrained is the determination of theheading offset between the direction a smartphone is facingand the direction the user is moving. There are approachesthat attempt to solve this problem [70]–[72], however it isstill an open research problem. Different transit modes suchas pedestrian, elevators and escalators make the sensor basedpositioning even more complicated. The approaches also haveto work on different types of surface and across wide group ofpeople regardless of their height, weight, age, gender, physicalcondition, etc.

Smartphone sensors can assist in pedestrian positioning byperforming the following tasks (Fig. 12): walking parametersestimation including step counting and step length and direc-tion estimation, motion classification, transit mode detection,

13

!

Fig. 13. Vertical acceleration during normal walking as it measured by atorso mounted accelerometer.

and floor change detection in multi-storey buildings. Thesetasks are carried out using a so called pedestrian dead reckon-ing (PDR) algorithm that computes position by integrating thedisplacement vectors, which represent steps. If no steps aredetected, the system is assumed to be stationary. The PDR’sperformance is quite robust to the sensor quality, which makesit suitable for implementation on smartphones.

Any PDR system performs the following tasks [73]:• Step count or step segmentation• Step length estimation• Step direction estimationIn the remainder of this section different approaches for

implementation of the algorithms shown in Fig. 12 will bedescribed: step detection, step length and direction estimation,orientation tracking, multiple usage scenarios and activityrecognition. Special attention will be paid to the methods thatdo not restrict a smartphone placement and usage: it may befirmly held in place, or free to move from its current positionwithout warning.

A. Walk detection and step counting

Most walk detection and step counting methods are based onthe assumption that pedestrian motion has a cyclic nature. Thealgorithms search for the repeating patterns in accelerometeror gyroscope data during walking (Fig. 13). They can begenerally categorized into the following groups: thresholdingand peak detection, auto-correlation, and spectral analysis. Thepros and cons of different walk detection and step countingapproaches are summarized in Table I.

1) Thresholding and peak detection: Thresholding andpeak detection algorithms are often used together: the walkingor running motion is detected applying thresholds and thensteps are counted by counting peaks [74]. In most cases stepsare detected and counted using accelerometers [75]–[83] and,in some cases, gyroscopes [84]. Simple algorithms apply athreshold on some parameters derived from the accelerometersignal. Walking and other activities are identified by compar-ison of different features with the pre-defined thresholds [74],[85]. The features are computed from windows of accelerome-ters and gyroscopes measurements, and usually include mean,standard deviation, energy, correlation, frequency-domain en-tropy, and FFT coefficients [74], [85], [86].

According to [86], a window length of 5-7 seconds issufficient to capture cycles in activities such as walking,running, window cleaning, or vacuuming. The mean and

energy of acceleration sufficed for accurate recognition ofthese activities. For computation of features in the frequencydomain a 512 sample window enables fast computation ofFFTs. The energy feature can be calculated in both time andfrequency domains. In the time-domain it is computed bysquaring the norm of the accelerometer and gyroscope dataand summing and normalizing them over a moving window.In the frequency domain it can be computed as the sum of thesquared FFT component magnitudes of the signal. The sum isdivided by the window length for normalization. Additionally,the DC component of the FFT should be excluded from thissum since the presence of a non-zero DC component canhide important information and reduce the effectiveness offrequency domain estimation techniques [85], [86].

Some algorithms use thresholds on vertical accelerationassuming that sensors are firmly attached and their orientationwith respect to the geographical frame is known. The per-formance of these algorithms can significantly deteriorate ifa smartphone is not firmly attached to the body. To resolvethe orientation problem the signal magnitude can be usedinstead. Another difficulty with the threshold approach is thatthe optimal threshold value can change significantly acrossdifferent users, surfaces and movements. This method maynot work correctly for unconstrained smartphones when themeasured acceleration does not match the regular accelerationpatterns during walking motion and the thresholds can betriggered mistakenly.

Another useful feature for walking motion detection is zero-crossing [75], [76], [78]. It is related to threshold methods andcan be computed based on accelerometer data. The decisionon whether a step is being made is done by analyzing themagnitude of the acceleration vector, which can be computedby subtracting the local gravity from the measured magnitudeof the specific force. When the sign changes from negative topositive, a new step is counted. The zero-crossing method isbased on the cyclic movement of the human body and requiresa triaxial accelerometer. Since the norm of the accelerationvector is used the orientation of the sensor unit has no effecton the measurement. This is a popular choice for pedometersor activity monitors due to its simplicity.

The steps can be counted by identifying and counting thepeaks (maximum or minimum detection) in the measuredacceleration. These peaks are emphasized in vertical accel-eration during walking because of heel impacts especiallywhen sensors are attached to the body [79], [80]. The peakscan be also detected in the norm of acceleration vector forother smartphone locations including texting or carrying ina bag. However, their magnitude is usually smaller. Thedifficulty with this approach is that each foot impact maygenerate multiple local peaks which can significantly increasethe algorithm complexity. The algorithm performance can beimproved if accelerometer measurements are pre-processed byapplying band-pass filtering. For example, Ladetto applied thewavelet transformation to remove high-frequency harmonicsfrom the signal [79]. A disadvantage of this approach is thatwalking motion detection may be triggered by any motion witha similar periodicity to walking.

14

2) Correlation based algorithms: An alternative to peak de-tection in acceleration signal is computing cross-correlation ofaccelerometer measurements with a pre-defined stride templateor auto-correlation function. The template can be previouslyrecorded walking data from the same person [73]. Correlationbased algorithms detect peaks in the mean-adjusted computedcorrelation of a temporal sequence of acceleration, usually thenorm of the measured acceleration vector. This approach canbe used for different sensor locations.

Rai et al. [83] proposed Normalized Auto-correlation basedStep Counting (NASC). They observed that if the user iswalking, then the auto-correlation will spike at the correctperiodicity of the walker. Given a sequence of accelerationsamples a(k), the normalized auto-correlation for lag τ at them-th sample is

χ(m, τ) =∑k=τ−1

k=0(a(m+k)−µ(m,τ))(a(m+k+τ)−µ(m+τ,τ))

τσ(m,τ)σ(m+τ,τ)

(17)

where the value of a time lag τ is chosen between 0.8 and2 sec. According to [83] when the person is walking and thestep frequency is equal to τ , the normalized auto-correlation isclose to one. Since, the value of τ is not known a priori, thealgorithm computes the auto-correlation for different valuesof τ within a pre-defined range and finds the value of τ forwhich the χ(m, τ) becomes maximum. NASC can also detectother repetitive activity such as running.

Dynamic time warping (DTW) algorithm attempts to rec-ognize each step within the data. A stride template is formedoffline and cross-correlation with the accelerometer signalis computed online [87]. DTW gives better results sinceit allows a non-linear mapping between the template andmeasured acceleration. However, good performance comes atthe price of a considerable computational overhead. Compareto DTW, computation of auto-correlation is faster and itscomputational burden can be further reduced by evaluating theauto-correlation only for a subset of time lags correspondingto the frequencies of interest [88].

3) Spectral analysis based algorithms: Another version ofpeak detection algorithm uses wavelet decomposition to ex-tract dominant frequency component from subsets of measuredacceleration [79] within the time-interval that includes at leasttwo cycles. In this approach the accelerometer signal is splitinto its low- and high-frequency components and the dominantfrequency taken as the walking frequency. By iterating thisprocess N times on the resulted low-frequency approximation,the dominant low-frequency signal with much smoother shapeis generated. The frequency of this signal corresponds towalking cadence. This is called the wavelet decompositiontree, and N is the number of computed levels. Ladetto [79]performed decomposition at level 4 using the Meyer waveletfunction. The decomposed at this level signal has only onemaximum within each cycle and therefore steps can be easilyidentified and counted.

Susi et al. [85], [89] applied the short-term Fourier trans-form (STFT) to evaluate the frequency content of successivewindows of accelerometer signal. The basic assumption of thistechnique is that any non-stationary signal can be regarded as

stationary for a short time duration. The walking cadence isvisible in the frequency peaks of the spectrogram obtained asthe square absolute value of the STFT. They also analyzedthe dominant frequency by evaluating the largest maximumin the spectrogram. When the user was walking and carryingthe inertial measurement unit (IMU) in the texting mode andwith the hand swinging, the walking cadence was visible inthe spectrogram’s frequency peaks and the strongest frequencyfor both IMU placements was coupled with the step event.Another observation was that when the user was performingan irregular motion, e.g. looking for a mobile phone in a bag,the two main frequencies showed a very irregular trend. Thespectrogram was completely different when the user was static:the dominant frequencies did not exhibit strong peaks. So, thefrequency analysis can be used to distinguish irregular motionmodes from normal walking [85].

Brajdic and Harle [74] used STFT for both walk detectionand step counting. In walk detection they used STFT to splitthe signal into successive time windows and labelled each aswalking if it contained significant spectral energy at typicalwalking frequencies. For step counting they used STFT tocompute a fractional number of strides for each window bydividing the window width by the dominant walking periodit detected. These fractional values were then summed toestimate the number of strides taken.

4) Walk detection and step counting on unconstrainedsmartphone: It should be noted that most implementationsclaim to use only the vertical acceleration, but do not com-pensate for changes in the global pose of the sensors duringwalking. Instead they assume that one of the accelerometeraxes remains vertical throughout. This assumption in generalis not valid for unconstrained smartphones. However, there aresome efficient and robust walk detection and step countingalgorithms designed for unconstrained smartphones. Theseapproaches can be classified into the following groups: (a)allowing arbitrary pose and carrying location, (b) tracking thephone orientation, and (c) performing motion classification andadapting different algorithms for different motions.

Pan et al. [90] proposed a step counting algorithm forunrestricted smartphone allowing arbitrary device pose duringwalking. The algorithm computes the phone pose transforma-tion based on gravity vectors. Then it looks for correlatedsegments from the collected raw data. The authors adoptedthe concept of correlation coefficients to identify whether thecollected sensing measurements exhibit similar tendencies andthen perform step segmentation. It was claimed that the stepdetection error is quite small when the phone is carried withoutchanging the carrying mode. This approach can be appliedonly for walking.

Susi et al. [85] analyzed biomechanics of walking and foundthat the swinging of the arm is synchronized with the footmotion, thus making walk detection and step counting possiblewhen sensors are carried in hand. The authors proposed to usethe gyroscope data for step counting because a swinging armproduces a notable cyclic pattern in angular velocity. So, thesteps can be counted by counting the peaks of the gyroscopesignal.

The same phenomena was also exploited in [91] where

15

TABLE ISUMMARY OF SMARTPHONE-BASED WALK DETECTION AND STEP COUNTING

Method Advantages DisadvantagesThreshold on magnitude of the acceler-ation

Simplicity, low computational burdenRobust to the phone’s orientation

Walking motion detection may be triggered by anymotion with a similar periodicity to walkingMay not work on unconstrained smartphones

Threshold on vertical acceleration Simplicity, low computational burden Requires accurate smartphone’s orientationMay not work on unconstrained smartphones

Threshold on features computed fromwindows of accelerometer and gyro-scope measurements

In addition to walking other activities can be identi-fied

Optimal threshold value can change significantlyacross different users, surfaces and movements

Peak detection for step counting Simplicity, low computational burden A foot impact may generate multiple local peaksMay not work on unconstrained smartphones

Normalized auto-correlation based stepcounting

Can be used for unconstrained smartphones Requires learning phase for each userTime lag 0.8-2 sec

Dynamic time warping Good performance in walk detection and step count-ing

High computational burdenA stride template is formed offline

Short-term Fourier transform The same algorithm works for walk detection andstep countingCan be used for unconstrained smartphonesCan distinguish irregular motion from walking

A window length of 5-7 seconds is required tocapture cycles in activities such as walking andrunning

Wavelet decomposition The same algorithm works for walk detection andstep countingCan be used for unconstrained smartphones

The data has to include at least two cycles

Renaudin et al. showed that the strongest frequency of thesignal extracted from the handheld IMU is coupled with thestep frequency of the walking person and can be used toestimate the cadence. The relationship between step and handfrequencies was analyzed for different hand motions and phonecarrying modes. The approach for reliable step frequencyestimation on handheld device is based on STFT. The fieldtests showed that the error in distance estimation with thisalgorithm was between 2.5–5% of the travelled distance [91],which is comparable with the accuracy achieved by otheralgorithms that are designed only for body fixed sensors.

Steinhoff and Schiele [92] experimented with a mediumaccuracy IMU (Xsens MTx) carried in trousers’ front pocketsduring walking. The main contribution of their work is ex-perimental evaluation of novel and existing PDR algorithmsin scenarios when an unconstrained IMU is carried in pocket.To cope with unknown and changing IMU pose the approachcomputes its global orientation. Then it computes the verticalacceleration, which was found to be a feature that is robustto changing orientation. To make the algorithm more robustthe signal was pre-processed with a median filter for noiseremoval. Also a threshold on the vertical acceleration variancewas applied to avoid spurious step detections from smallpeaks. Based on a field test with eight people it was foundthat the step detection accuracy was not affected by the IMUposition inside the pocket or the shape of the pocket.

Summarizing the walk detection and step counting algo-rithms for unconstrained smartphones it can be concludedthat all successful algorithms are able to recognize (or knowa priori) how a smartphone is being carried and then adaptthe algorithms for better performance. Some approaches evendemonstrated the ability to cope with ill-defined movementssuch as shuffling or side steps [74].

5) Step counting during walking on stairs or ramps:Another common assumption in step detection is that mostalgorithms are developed and tested only for pedestrian motionon flat surfaces. Ladetto [79] tested step counting algorithms

on ramps and found that most algorithms are robust to smallslopes in surface and start to break down only on inclines of10% or more. More recently, Wang et al. [93] have demon-strated that different gait patterns corresponding to differentinclines can be distinguished autonomously with accuracyexceeding 90%. From this we can conclude that algorithmsfor walk detection and step counting can be adapted to copewith long ramps such as those for wheelchair access, whichcan be found in many buildings.

Munoz et. al [84] proposed step detection algorithms fora pocket-mounted sensor unit that can work not only duringwalking on flat surfaces, but also during climbing and de-scending stairs. This algorithm exploits the use of the pitchangle computed by the IMU. In addition to step segmentationit was demonstrated that using the pitch angle, it is possibleto identify five basic physical activities: standing, sitting,walking, walking up and down stairs. However, this algorithmheavily relies on the assumption that the sensor unit is in apocket. When the unit is taken out of the pocket the algorithmmay perform incorrectly.

6) Performance of step counting algorithms implementedon smartphones: The most comprehensive presentation offield test results for walk detection and step counting onunconstrained smartphones is given in [74]. The authorsproposed metrics and scenarios that allow uniform comparisonof algorithms on commodity smartphones. The test scenariosincluded different activities such as picking up and puttingdown the phone, texting, revolving on an office chair, transi-tioning between standing and sitting, and between walking andstanding. Step counting errors are computed for eight differentalgorithms based on a set of 130 sensor traces from 27 distinctusers. All algorithms require only acceleration magnitude datagathered from a smartphone.

The results for walk detection showed that the best perfor-mance was achieved by the two threshold based algorithms:one with the threshold on standard deviation and anotheron signal energy. They exhibited total median error in walk

16

detection of about 1% including both false positive and falsenegative errors. The overall best step counting results wereobtained using the windowed peak detection algorithm witha median error of 1.3%. Except for the placement in a backtrouser pocket, the majority of placements have little effect onthe step counting performance.

The threshold based algorithms can have false-positive de-tection errors, reporting steps even when users are not walkingif they make sudden movements such as hand gestures, turningor shifting in the chair etc. When walking is disrupted bysudden movements Brajdic and Harle [74] used normalizedauto-correlation algorithm combined with the threshold onstandard deviation. This approach is more robust since itexploits the repetitive nature of walking.

If the phone is placed in a trouser pocket the step countingcan be improved by adding gyroscopes data to the algorithm.Munoz et al. [84] presented a step detector and step lengthestimator based on the opening angle of the leg. It was shownthat the movement of the user’s leg is strongly correlated withwalking cadence and the step length. This approach can alsoreliably work during climbing and descending stairs. The errorin step detection is less then 1%, however, the field tests werecarried out with Xsens IMU, which has much more accurategyroscopes than any smartphone.

B. Step Length Estimation

Simple algorithms only count steps assuming that the steplength is just the average for that user. Advanced algorithmsalso perform accurate step segmentation and analyze theaccelerometer signals to estimate the magnitude of each stepindividually. However, most of these systems require calibra-tion to an individual user because everyone’s gait has differentacceleration profiles.

The following two formulas can be used to estimate steplength when the smartphone is in a jacket or trouser pocket.Since the formulas use vertical component of the accelerationto compute a step length, the mobile device’s pose has tobe known. An empirical relation between maximum andminimum of the vertical acceleration and the step length isgiven by [77]

SL1 = K 4√amax − amin (18)

where amax, amin are the maximum and minimum values of thevertical acceleration during one step, and K is a calibrationconstant that can adjust step length estimation to a specificperson. This formula is based on the bounce movement of thehip while walking. The second step length estimator that wasdeveloped for implementation on smartphones when they arecarried in pockets [94], [95] is

SL2 = 0.1 2.7

√√√√∑k=Ni=1 ‖avert,i‖

N

√K√

T · apeak(19)

where T is a step duration, avert,i is a sequence of measuredvertical accelerations during one step, apeak is the differencebetween maximum and minimum vertical acceleration duringthe same step, and K is a calibration constant.

Ladetto [79] proposed an empirical formula for step lengththat uses the variance of acceleration instead of the verticalcomponent of acceleration. His formula is the following rela-tion between the step length, step frequency and variance ofacceleration signal:

SL3 = K1 +K2 f +K3 Var + w (20)

where K1,K2,K3 are user-specific calibration parameters, fis an estimated step frequency, Var is the variance of themeasured acceleration signal, and w is a Gaussian noise.The accuracy of this formula was analyzed during field testswith 20 persons who walked with a constant cadence. Theresults indicate that the step length computed by Eq. 20 isquite accurate for fast walking with the variance of 4% at130 steps/min. However, the variance increased to 15% at acadence of 60 steps/min. Fortunately, since the mean error ismuch smaller the error in traveled distance is less than 2%.

Gusenbauer et al. [96] also used a linear equation to modelthe relation between the step length and cadence, but withoutthe acceleration variance term:

SL4 = K1 +K2 f + w (21)

They regarded the step frequency as a crucial parameter forthe step length estimation. The natural asymmetry in left andright steps is accounted for by a Gaussian noise term w. Basedon field tests with different subjects, the standard deviation forthe Gaussian noise distribution was defined as

σ = 0.24− 0.09f (22)

A model for reliable step length estimation on handhelddevices was proposed in [91]. It includes the step frequency,height, and three calibration parameters:

SL5 = h · (K1 · f +K2) +K3 (23)

where h is the user’s height. The authors proposed to usetwo sets of parameters: universal and calibrated. The universalparameters are not calibrated to a specific person and usuallygive the first approximation for the calibration procedure.The calibrated parameters are computed using recursive least-squares by minimizing the sum of squared residuals betweenthe true step length and the predicted step length for nparticular people. The relationship between walking cadenceand hand swinging frequency was analyzed for different handmotions and sensor carrying modes based on the frequencycontent of synchronized signals collected with two sensorsplaced in the hand and on the foot.

The approach proposed in [84] works only when the mobiledevice is carried in trouser pockets during walking and itcomputes the step length using the opening angle of the leg.Based on extensive field tests the authors found that thisrelationship can be applied in general for all users regardless ofother parameters such as the user’s height, leg length, weight,gender and the pocket’s size. The step length is computedusing the formula

SL6 = K1 +K2 ∆θ (24)

where ∆θ is the pitch angle measured by an IMU in trouserpocket, and K1,K2 are user-specific calibration parameters.

17

The results showed that the estimation of the step length isinsensitive to speed changes.

1) Performance of step length estimation algorithms withunconstrained IMU: Jahn et al. [95] evaluated the performanceof the step length estimation algorithms based on Eq. 18 andEq. 19. In the first case the Xsens IMU was attached to theback approximately at the center of mass and in the secondcase the IMU was carried in a trouser pocket. In both cases theerror in step length estimation was about 2–3%. The standarddeviation in the second case was higher. The measurement datawas collected from five participants during indoor field testson a flat concrete area. The subjects were asked to completethe test course at three different speeds: "normal", "slow" and"fast". The calibration constant was computed based on testdata for different walking speeds.

Fang et al. [80] tested the step length estimator based onEq. 18 on a 1.2 km course in indoor and outdoor environmenton different terrains including swamp terrain. The overalldistance estimation error with personal calibration was about±3% and ±8% without calibration across the variety ofsubjects. The authors noticed that the estimation accuracydegrades rapidly when a person is walking, for example, onnon-flat terrain or climbing on stairs.

Ladetto [79] tested his approach with a body mounted IMUduring experiments with 20 people. He could achieve an errorin distance estimation of about 1% with parameters calibratedto each person. The tests also revealed that slopes smaller than10% have little effect on distance estimation.

Renaudin et al. [91] evaluated performance of the steplength estimation algorithm with hand-held IMU during fieldtests involving 10 test subjects. They also compared theaccuracy of travelled distance estimation of the same algorithmbetween the case when universal parameters were used andwhen a set of parameters was calibrated for each subject.The algorithm with user-specific parameters showed an errorbetween 2.5 and 5% of the travelled distance, which iscomparable with that achieved by models proposed in theliterature for body fixed sensors only. The algorithm withuniversal parameters showed an error between 3.2 and 9%of the travelled distance.

C. Walking direction estimation

Walking direction estimation on an unconstrained smart-phone is the most difficult problem that has to be solvedto achieve capabilities for autonomous navigation. This isstill an open problem since most of the proposed solutionscan work only with IMUs that have much better accuracythan modern smartphone sensors. The algorithms for walkingdirection estimation usually include two important tasks: (a)computation of the smartphone’s pose with respect to someglobal frame; (b) estimation of walking direction in the globalframe using the smartphone’s sensors.

1) Computing smartphone’s orientation in global frame:The smartphone’s orientation is usually computed with respectto a global frame N as shown in Fig. 14. This frame isfixed to the earth and its xn and yn axes are aligned withthe local horizontal plane pointing north and east respectively,

!!!

gravity vector

horizontal plane xn!

yn!

zn!

xb!

yb!

zb!

smartphone !

Fig. 14. Relation between global and smartphone’s body (sensor) frames.Rb

n is the rotation matrix from body frame to global frame.

and the zn axis is aligned with the local vertical. This way anorthogonal right-hand coordinate frame is formed, sometimesreferred to as east-north-up (ENU) frame. The sensor or bodyframe B is centered in the smartphone and moves with it.Ideally, its origin would be located in the accelerometer toavoid measuring accelerations due to pure rotation. However,since the accelerometer’s exact location is unknown, the originis assumed to be in the phone’s geometrical center. If thescreen is facing up, the x-axis points right, the y-axis pointsforward, and the z-axis points up. Both the Android and theiOS APIs use the same definition of the coordinate frames.Mathematically, orientation of the sensor frame with respect tothe global frame can be described by a transformation matrixRb

n which is composed of the direction cosine values betweenthe two coordinate frames.

Tracking orientation using smartphone’s gyroscopes is adifficult task that requires frequent updates from other aid-ing sources, for example, magnetometers and accelerometers.However, magnetometers are not reliable indoors because ofmagnetic field distortions and they cannot track fast move-ments like those when a phone is swinging with the hand. Onesolution to this problem is based on the assumption that duringwalking and some other activities the average accelerationdue to the movement is almost zero. Then, the direction ofgravity can be measured by the accelerometers and used tocorrect the horizontal angles. This assumption is valid for mostpractical cases encountered in pedestrian navigation, sincethe acceleration of a person’s body during walking is cyclic,and therefore, if it is averaged over one or several cycles itwill be almost zero. The accuracy of the gyroscopes must begood enough to track the orientation during the time betweenattitude updates to a high degree of accuracy. Since the gyrosin smartphones are not accurate this time must be very short,only 1-2 seconds.

The algorithm for pitch and roll angles computation on asmartphone based on the complementary filtering techniqueand quaternion mechanization was proposed in [70], [97].In the filter prediction stage the state mean and covarianceare propagated using the gyroscope measurements, while themeasurements from the accelerometers and magnetometersare used during the update stage. The approach relies on thecapability to calculate and resolve the absolute instantaneousroll and pitch angles of the device by using measurements fromaccelerometers. The data is first low-pass filtered or averagedto reduce noise and then the roll and pitch are computed using

18

the equations [70]:

roll = arctan(−fx/fz)pitch = arctan(fy/

√f2x + f2z )

(25)

where fx, fy, fz are the instantaneous or averaged accelerationmeasurements along x, y and z axes. Averaging accelerometersignals reduces not only noise, but also the sensor’s bandwidth.As a consequence, the averaged accelerometer signals mainlymeasure the gravity components, which leads to more accuratecomputation of pitch and roll. In real-time implementationsthe noise can be also reduced by a moving average methodeither applied to the calculated roll and pitch or applied tothe accelerometer readings themselves before the calculationof roll and pitch [70].

In [98] another approach for horizontal angles estimationwas proposed based on low-cost IMU and external velocityaiding obtained from a model of human gait kinematics, whichprovides an alternative way to calculate traveled distanceand velocity averaged during the step. The algorithm wasapplied to pedestrian navigation systems that included a torsomounted IMU. The kinetic model of gait is considered as avirtual sensor that can be used to estimate the step eventsand step size for a person while walking. The term "virtual"emphasizes the fact that there is no separate instrument fordirect speed measurement. The speed is estimated using ac-celerometer measurements and additional information aboutthe pedestrian’s movement which is derived from the model ofhuman gait kinematics. The different characteristics of errorsin INS output and in this virtual measurement make it possibleto apply complementary filter methodology and significantlyimprove INS performance by keeping the horizontal velocityand tilt errors small.

After the horizontal angles are computed the azimuth canbe determined. If a smartphone is being used, combining agyro and a magnetic compass will give better results thanusing these sensors separately. The magnetic compass is ableto provide the external azimuth to update the parameters of thegyro and the gyro can be used to detect the disturbances thatcan be as large as tens of degrees. Ladetto [99] proposed usinga Kalman filter for coupling a magnetic compass with a low-cost gyroscope. In this case, the strengths of one technologycan compensate the weaknesses of the other. If we comparethe rate of change of both signals while measuring the strengthof the magnetic field, it is possible to detect and compensatemagnetic disturbances. In the absence of such disturbances,the continuous measurement of the azimuth allows to estimateand compensate the bias and the scale factor of the gyroscope.The reliability of indoor and outdoor navigation improvessignificantly thanks to the redundancy in the information.

2) Walking direction estimation: The ability to accuratelyestimate walking direction, regardless of smartphone orien-tation, is extremely important for long term positioning. Thedifficulty in walking direction computation with unconstrainedsmartphones comes from the fact that the direction that asmartphone is facing is often different from the direction theuser is moving. So, even if the smartphone’s pose with respectto a global frame is accurately determined this is not enoughto estimate the step direction.

Traditional IMU can determine the direction of travel bycomputing vectors of acceleration or velocity increments inthe global frame, if the horizontal angles are accuratelycomputed. This approach is usually not possible to implementwith smartphone sensors. Other methods for walking directionestimation are focused on estimation of the misalignmentbetween a smartphone and the pedestrian, or in other wordsthe misalignment between the frame of the sensor assembly inthe device and the frame of the pedestrian [100], or walkingdirection [72]. Algorithms for angular misalignment estimationbetween the smartphones pointing direction and pedestrianwalking direction include [72]:

1) The principal component analysis method (PCA)2) The forward and lateral accelerations modeling (FLAM)3) The frequency analysis of inertial signals (FIS)

The first two algorithms require only the acceleration mea-sured by a smartphone. The frequency analysis method alsorequires the angular velocity in addition to the acceleration.These algorithms can be potentially implemented on smart-phones in different carrying modes including handheld.

The prerequisite for the above mentioned methods is com-putation of the smartphone’s pose in the global frame, sincethe measured acceleration in the sensor frame has to be trans-formed to the global frame. In the PCA method the first step isto compute horizontal acceleration components and remove thegravity value from the acceleration. Then the two horizontalacceleration components are fed to the PCA algorithm, whichgenerates the principal components based on two features ofhuman walk [72], [92], [100]–[102]: (a) over a gait cyclethe variance of the horizontal acceleration is maximum alongthe forward direction; (b) the horizontal acceleration varianceis minimum along the lateral direction. The directions arecomputed using eigenvalues: the eigenvector correspondingto the largest eigenvalue gives the forward direction, and thesecond eigenvalue’s eigenvector gives the lateral direction.

The FLAM method for estimating the angular misalignmentis based on modelling of the forward and lateral accelerationsby a sum of sinusoids and computing the correlation functionfor the modelled and measured acceleration. When the angleparameter of this function is equal to the walking direction thecorrelation function reaches its maximum [103].

The FIS method also exploits the fact that the horizon-tal acceleration is maximum along the forward direction. Ituses a cost function containing the power functions of theacceleration and angular velocity in the forward and lateraldirections. These power functions are computed using a short-time Fourier transform. The walking direction is estimated bymaximizing this cost function over the angle parameter [72].

The uDirect approach [104] determines a user’s facingdirection with a mobile phone, regardless of the mobilephone’s orientation. A user’s forward direction is computedbased on accelerometers, exploiting the physiological walkingbehavior of a human. By analyzing the acceleration pattern ofthe thigh movement during walking locomotion, the algorithmis able to identify the optimum moment where the capturedacceleration signal mainly consists of the forward direction,minimizing the noise of unwanted side components. The un-derlying mathematical model using the quaternion presentation

19

further simplifies computation of the rotation compared to avector based approach.

3) Performance of the walking direction estimation algo-rithms implemented on smartphones: The Humaine [101]positioning system was implemented on Android devices. Thefield tests included 170 experiments on different testbeds andphone positions and showed that the system works accuratelyindoors and outdoors, without user involvement or prior con-figurations. Humaine achieved 15 degrees of median accuracyin walking direction estimation. The uDirect [104] autonomouspositioning system was tested only with a smartphone carriedin trouser pocket and it could estimate walking direction with22 degrees of error.

Other algorithms for walking direction estimation havebeen tested only with tactical grade IMUs that included atri-axial accelerometer, a tri-axial gyroscope and a tri-axialmagnetometer: ADIS-16488 in [72], ADIS-16407 in [71],Xsens MTx in [92], [102], etc. The reported walking directionestimation accuracy for these algorithms with IMU carried inhand or in trouser pockets is about 3 − 5 deg. However, itis not possible to predict what these algorithms’ accuracy inwalking direction estimation would be if the high quality IMUwere replaced by a smartphone.

D. Multiple usage scenarios and transit modes

Identification of transit mode, type of activity and phoneplacement can significantly improve the accuracy and relia-bility of PDR and map-matching. For typical smartphone de-ployment scenarios it will be enough to identify the followingmodes of transit and activities [74], [83], [105]: idle, pickingup and putting down the phone, texting, transitioning betweenstanding and sitting, level walking, up- and downstairs climb-ing, riding on elevators and escalators. In addition to theseactivities some common phone placement scenarios have tobe considered: hand-held, front trouser pocket, back trouserpocket and in a backpack or handbag. All movements can beperformed at different speeds and intensity levels.

1) Motion classes: For the purpose of pedestrian navigationthe motion classification can be simplified by limiting thenumber of classes to only five, similar to what was proposedin [85]:

1) All situations when both the device and the user arestatic, i.e. the location does not change significantlyduring some time with the allowed amount of movementnot exceeding the predefined threshold.

2) All cases when a user is walking except when a phone isheld in swinging hand. This class includes the followingcases: (a) A phone is in pocket, bag, arm-band or anyother type of attachment; (b) Texting or watching screen;(c) Making or receiving a call.

3) All cases when the user is walking while holding themobile device in his/her swinging hand.

4) Moving on elevators or escalators.5) Irregular motions that the user performs while standing

or sitting. During this time the user’s location does notchange.

2) Feature sets: Time domain features include basic signalcharacteristics and statistics. The following features can beused to identify different types of activities for pedestriannavigation: magnitude, variance and energy of accelerometerand gyroscope measurements [74], [85]. Frequency-domainfeatures can be used to capture periodicity in data and dis-tinguish cyclic activities from irregular movements. They canbe computed using coefficients of Fourier transform or fastFourier transform (FFT), which also provide the dominantfrequencies in the measured acceleration and angular velocity.

Since the bandwidth of human movements is usually below10 Hz, the sensor data can be low-pass filtered, for example,using a 10th order Butterworth filter with a 15 Hz cut-offfrequency [85]. Filtering can improve motion classificationbecause the smartphone sensors are noisy. During fast move-ments, for example, when a smartphone is swinging with thehand, the norm of the acceleration and angular velocity canbe considered. This removes dependence on the phone poseand makes motion classification more robust.

3) Motion classifier: Pedestrian motion can be classified bya decision tree, which uses a flowchart-like graph of decisionsand their possible consequences covering all of the possibleclasses. Each internal node represents comparison of featureswith the thresholds, each branch represents the outcome ofthis comparison and each end node represents a class label.The paths from root to leaf represents classification rules. Adecision tree consists of 3 types of nodes: decision nodes,chance nodes, end nodes.

In [85], [106] the authors used a multivariate decision tree.At the first node the algorithm evaluates the energies andvariances of gyroscope and accelerometer signals to determinestatic and dynamic activities. Then the features describing peri-odicity are checked to separate irregular motion from walking.Irregular motion is usually characterized by very high valuesof the gyroscope and accelerometer variances in short temporalperiods. Finally, the gyroscope measurements are evaluated toidentify walking with a swinging hand, which is characterizedby high values of the gyroscope and accelerometer variancesand energies, and frequency peaks in the STFT spectrogramof the gyroscope signal.

E. Summary of sensor based positioning

Although in contemporary smartphones the sensor basedpositioning cannot operate as a stand-alone technology itcan be very useful when combined with WLAN/BLE basedpositioning or map-matching [107]. The major advantage ofthis technology is availability of accelerometers, gyroscopesand magnetometers in almost every smartphone. The majordifficulty in implementation of these tasks on smartphonesis freedom of movement and arbitrary placement. The majortasks that can be performed using smartphone sensors are stepcount, distance and walking direction computation, activityrecognition, vertical movements and floor change detection.The most difficult task is a smartphone orientation trackingand walking direction estimation.

Examples of autonomous positioning implemented on atypical Android smartphone are shown in [108], [109]. During

20

the test, the user was carrying the phone in hand and pocket,texting and making a call with the phone on the ear. Duringtwo minutes of autonomous navigation the position error didnot exceed 7 meters. In another test the user was walkingindoors for 4 minutes and the maximum error stayed within 13meters, even with the phone changing orientation with respectto the user. This fully autonomous solution can be furtherimproved when combined with WLAN measurements and thebuilding floor plan.

VI. OPEN RESEARCH ISSUES

Smartphone based indoor positioning systems are being de-ployed on Apple and Google smartphones and provide reason-ably good positioning accuracy in areas rich with WLAN APsand BLE beacons. The companies also provide a public APIfor Wi-Fi, BLE, inertial and magnetic sensors data that canbe used by third party developers. However, the positioningperformance still can be improved and open research issuesinclude autonomous indoor positioning, integrity monitoring,fingerprint database generation and maintenance, seamlesspositioning between indoors, outdoors and different modesof transit, evaluation and comparison of different positioningalgorithms

1) Autonomous indoor positioning: Currently, there is noeffective approach for long-term autonomous indoor position-ing when absolute position fixes from WLAN or BLE areunavailable. A possible solution can be based on pedestriandead-reckoning (PDR) combined with maps. However, PDRimplementation for unconstrained phones and different modesof transit is complicated.

2) Integrity monitoring: Receiver autonomous integritymonitoring is commonly used in GNSS. There are also someapproaches for integrity monitoring of INS/GNSS integratednavigation systems. In indoor positioning integrity monitoringmakes it possible to improve operational robustness and elim-inate outliers, in particular, to isolate WLAN APs and BLEbeacons that disagree with the fingerprint database. There is aneed for such methods on smartphone based indoor positioningsystems. Possible solutions can use hardware and analyticalredundancy: multiple WLAN APs and BLE beacons, PDR,map-matching and magnetic field fingerprinting.

3) Effective generation and maintenance of fingerprintdatabase: Since the data collection required for WLAN andBLE fingerprint database generation is a time consuming andlabor intensive process, the crowdsourcing approach will playa critical role in reducing time and costs in the future. Themajor difficulty is in providing reference position indoors.

4) Positioning performance evaluation: There is no stan-dard procedure for positioning accuracy evaluation of differentalgorithms. Most publications present only walking tests fora single user in a small area of an office building filled withWLAN APs and BLE beacons. The tests are performed in avery controlled environment in which the same device and theway of carrying is used during the training and positioningphases. In some cases the reported position accuracy can be1-2 m. However, in real life situations the position accuracyis usually significantly worse, especially when the fingerprint

database is not up-to-date or the WLAN and BLE coverage isnot complete or the fingerprints are not collected in all areas.The accuracy can degrade also because people use differenttypes of phones and carrying modes.

A possible solution to this problem can be creation ofa public database that can be used as a benchmark duringevaluation of different indoor positioning algorithms. Thegeneral requirements to this database can be summarized as

• The data has to be recorded on several different com-modity smartphones and include RSS measurements fromall available WLAN APs and BLE beacons, acceleration,angular velocity, magnetic field, GNSS or assisted GNSS.

• The WLAN and BLE fingerprints as well as the referencesmartphone position has to be provided.

• The test path has to be long enough to include differenttypes of buildings such as office buildings, shoppingmalls, convention centers, airports, etc, floor change inmulti-storey buildings, using elevators and escalators, anddifferent modes of transit between the buildings (car, bus,subway, and train).

• The data is recorded during field tests with many differentusers who follow the same pre-defined scenario thatincludes carrying the phone in hand and pocket, textingand making a call with the phone on the ear.

The work presented in [43] is a step towards such a publicbenchmark.

VII. CONCLUSIONS

In this paper we have surveyed the most significant algo-rithms that can be implemented in smartphone positioningsystems: WLAN/BLE based positioning, map-matching, mag-netic field fingerprinting, and motion tracking using sensors.The default approach for indoor navigation on a smartphone isbased on WLAN. The positioning accuracy depends on densityof WLAN access points and calibration points, and can besufficient for many applications in typical public buildings likeshopping malls, universities, airports etc. The disadvantage ofthis approach is the labor-intensive process of building andmaintaining the WLAN database.

Improvement of WLAN based positioning can be achievedby using indoor map information. The assumption of indoormap availability is justified since the map is necessary tocommunicate results to the user and also for creating the radiomap. So, we can conclude that map aiding does not requireany additional investments in hardware or infrastructure. Thiscapability is enabled by software that implements a map-matching algorithm.

Many modern smartphones also contain accelerometers,gyroscopes and magnetometers. The use of sensors can sig-nificantly enhance the positioning accuracy of WLAN/BLEbased solution because the advantages and disadvantagesof the two systems broadly complement each other: sensorbased position errors growth can be curbed using WLANpositions and gaps in WLAN/BLE coverage can be filledby sensor based positioning. However, because smartphonesensors are not accurate, long term autonomous positioningwithout WLAN/BLE position fixes currently is not possible.

21

Walking direction computation is the biggest challenge inimplementation of PDR on smartphones and it is still an openresearch problem.

Magnetic field fingerprinting can be added to other indoorpositioning methods when high position accuracy is required.Although this method can potentially provide sub-meter posi-tion accuracy its implementation on smartphones is difficult,mainly because of stringent requirements to magnetic mapdatabase. Magnetic field gradients can be very steep indoorsand, therefore, a dense grid of calibration points is necessary.Since the magnetic field may also vary significantly withheight, the calibration points also have to be at differentheights. Methods have to be developed to create the databasein an efficient manner. Finally, transient magnetic interferencesshould be considered, especially in areas close to elevators.

The most accurate navigation solution is based on a combi-nation of WLAN, BLE, map, magnetic field and sensor basedpositioning. The trend for augmentation of WLAN/BLE basedpositioning by all possible sensors will continue in the futuresmartphone indoor positioning systems. In addition to WLANand BLE absolute position fixes can be also obtained fromLTE, 5G or other future technologies. These measurementswill be fed into data fusion filter that can combine them withmap, magnetic field fingerprints and sensor measurements andproduce accurate and reliable position estimates.

ACKNOWLEDGMENT

The authors thank Laura Wirola, Dr Helena Leppäkoski,Dr Philipp Müller, Henri Nurminen, Ville Honkavirta, andDr Simo Ali-Löytty for permission to use their material. Theauthors would also like to thank the reviewers of this paperfor their detailed and thoughtful comments that considerablyimproved this work.

REFERENCES

[1] IEEE Computer Society LAN/MAN Standards Committee, “WirelessLAN medium access control (MAC) and physical layer (PHY) speci-fications,” 1997.

[2] V. Honkavirta, Location fingerprinting methods in wireless local areanetworks. MSc Thesis, Tampere University of Technology, 2008.

[3] V. Honkavirta, T. Perälä, S. Ali-Löytty, and R. Piché, “A comparativesurvey of WLAN location fingerprinting methods,” in Proc. of the6th Workshop on Positioning, Navigation and Communication 2009.WPNC’09, Mar 2009, pp. 243–251.

[4] S. Ali-Löytty, T. Perälä, V. Honkavirta, and R. Piché, “FingerprintKalman filter in indoor positioning applications,” in Control Appli-cations, (CCA) Intelligent Control, (ISIC), 2009 IEEE, July 2009, pp.1678–1683.

[5] F. Seco, A. Jimenez, C. Prieto, J. Roa, and K. Koutsou, “A surveyof mathematical methods for indoor localization,” Intelligent SignalProcessing, pp. 9–14, 2009.

[6] H. Liu, H. Darabi, P. Banerjee, and J. Liu, “Survey of wireless indoorpositioning techniques and systems,” Systems, Man, and Cybernetics,Part C: Applications and Reviews, IEEE Transactions on, vol. 37, no. 6,pp. 1067–1080, 2007.

[7] P. Müller, M. Raitoharju, and R. Piché, “A field test of parametricWLAN-fingerprint-positioning methods,” in Proc. 17th Int Conf onInformation Fusion, Salamanca, Spain, July 2014.

[8] T. Roos, P. Myllymäki, H. Tirri, P. Misikangas, and J. Sievänen, “Aprobabilistic approach to WLAN user location estimation,” Interna-tional Journal of Wireless Information Networks, vol. 9, no. 3, pp.155–164, 2002.

[9] M. Youssef and A. Agrawala, “The horus WLAN location determi-nation system,” in Proceedings of the 3rd international conference onMobile systems, applications, and services. ACM, 2005, pp. 205–218.

[10] L. Koski, R. Piché, V. Kaseva, S. Ali-Löytty, and M. Hännikäinen,“Positioning with coverage area estimates generated from locationfingerprints,” in Proc. of the 7th Workshop on Positioning, Navigationand Communication. WPNC’10, Mar 2010, pp. 99–106.

[11] L. Koski, R. Piché, and S. Ali-Löytty, “Positioning with Bayesiancoverage area estimates and location fingerprints,” in Proc. EuropeanConference on Mathematics in Industry, Wuppertal, Germany, 2010.

[12] L. Koski, T. Perälä, and R. Piché, “Indoor positioning using WLANcoverage area estimates,” in Proc. Int. Conf. Indoor Positioning andIndoor Navigation, Sep 2010, pp. 1–7.

[13] H. Nurminen, Position estimation using RSS measurements with un-known measurement model parameters. MSc Thesis, Tampere Uni-versity of Technology, 2012.

[14] H. Nurminen, J. Talvitie, S. Ali-Löytty, P. Müller, E. Lohan, R. Piché,and M. Renfors, “Statistical path loss parameter estimation and posi-tioning using RSS measurements in indoor wireless network,” in Proc.Int. Conf. Indoor Positioning and Indoor Navigation, Nov 2012.

[15] C.-L. Wu, L.-C. Fu, and F.-L. Lian, “WLAN location determination ine-home via support vector classification,” in Networking, sensing andcontrol, 2004 IEEE international conference on, vol. 2. IEEE, 2004,pp. 1026–1031.

[16] YuFeng, J. Minghua, L. Jing, Q. Xiao, H. Ming, P. Tao, and H. Xinrong,“An improved indoor localization of WiFi based on support vectormachines,” International Journal of Future Generation Communicationand Networking, vol. 7, no. 5, pp. 191–206, 2014.

[17] D. Sánchez-Rodríguez, P. Hernández-Morera, J. M. Quinteiro, andI. Alonso-González, “A low complexity system based on multipleweighted decision trees for indoor localization,” Sensors, vol. 15, no. 6,pp. 14 809–14 829, 2015.

[18] C. Feng, W. S. A. Au, S. Valaee, and Z. Tan, “Received-signal-strength-based indoor positioning using compressive sensing,” IEEETransactions on Mobile Computing, vol. 11, no. 12, pp. 1983–1993,2012.

[19] B. Viel and M. Asplund, “Why is fingerprint-based indoor localizationstill so hard?” in Pervasive Computing and Communications Work-shops (PERCOM Workshops), 2014 IEEE International Conference on.IEEE, 2014, pp. 443–448.

[20] S. He and S.-H. G. Chan, “Wi-Fi fingerprint-based indoor positioning:Recent advances and comparisons,” IEEE Communications Surveys &Tutorials, vol. 18, no. 1, pp. 466–490, 2016.

[21] A. M. Hossain, Y. Jin, W.-S. Soh, and H. N. Van, “SSD: A robust RFlocation fingerprint addressing mobile devices’ heterogeneity,” IEEETransactions on Mobile Computing, vol. 12, no. 1, pp. 65–77, 2013.

[22] C. Laoudias, D. Zeinalipour-Yazti, and C. G. Panayiotou, “Crowd-sourced indoor localization for diverse devices through radiomapfusion,” in Indoor Positioning and Indoor Navigation (IPIN), 2013International Conference on. IEEE, 2013, pp. 1–7.

[23] Y. Chapre, P. Mohapatra, S. Jha, and A. Seneviratne, “Receivedsignal strength indicator and its analysis in a typical WLAN system(short paper),” in Local Computer Networks (LCN), 2013 IEEE 38thConference on. IEEE, 2013, pp. 304–307.

[24] F. Della Rosa, L. Xu, J. Nurmi, C. Laoudias, M. Pelosi, and A. Ter-rezza, “Hand-grip and body-loss impact on RSS measurements forlocalization of mass market devices,” in 2011 International Conferenceon Localization and GNSS (ICL-GNSS), 2011.

[25] F. Della Rosa, M. Pelosi, and J. Nurmi, “Human-induced effectson rss ranging measurements for cooperative positioning,” HindawiInternational Journal of Navigation and Observation, 2012.

[26] M. A. Youssef, A. Agrawala, and A. U. Shankar, “WLAN locationdetermination via clustering and probability distributions,” in PervasiveComputing and Communications, 2003.(PerCom 2003). Proceedings ofthe First IEEE International Conference on. IEEE, 2003, pp. 143–150.

[27] S.-H. Fang and T. Lin, “Principal component localization in indoorWLAN environments,” IEEE Transactions on Mobile Computing,vol. 11, no. 1, pp. 100–110, 2012.

[28] Y. Chapre, A. Ignjatovic, A. Seneviratne, and S. Jha, “CSI-MIMO:Indoor Wi-Fi fingerprinting system,” in 39th Annual IEEE Conferenceon Local Computer Networks. IEEE, 2014, pp. 202–209.

[29] S. Sen, J. Lee, K.-H. Kim, and P. Congdon, “Avoiding multipath torevive inbuilding WiFi localization,” in Proceeding of the 11th annualinternational conference on Mobile systems, applications, and services.ACM, 2013, pp. 249–262.

[30] D. Chen, L. Du, Z. Jiang, W. Xi, J. Han, K. Zhao, J. Zhao, Z. Wang,and R. Li, “A fine-grained indoor localization using multidimensionalWi-Fi fingerprinting,” in 2014 20th IEEE International Conference onParallel and Distributed Systems (ICPADS). IEEE, 2014, pp. 494–501.

22

[31] X. Wang, L. Gao, S. Mao, and S. Pandey, “DeepFi: Deep learningfor indoor fingerprinting using channel state information,” in 2015IEEE Wireless Communications and Networking Conference (WCNC).IEEE, 2015, pp. 1666–1671.

[32] X. Wang, L. Gao, and S. Mao, “PhaseFi: Phase fingerprinting for indoorlocalization with a deep learning approach,” in 2015 IEEE GlobalCommunications Conference (GLOBECOM). IEEE, 2015, pp. 1–6.

[33] R. Faragher and R. Harle, “An analysis of the accuracy of Bluetoothlow energy for indoor positioning applications,” in Proc. of the 27thInt. Technical Meeting of The Satellite Division of the Institute ofNavigation, Tampa, Florida. ION, Sep 2014, pp. 201–210.

[34] S. Ashok and R. Krishnaiah, “Overview and evaluation of bluetooth lowenergy: An emerging low-power wireless technology,” InternationalJournal, vol. 3, no. 9, 2013.

[35] X. Zhao, Z. Xiao, A. Markham, N. Trigoni, and Y. Ren, “DoesBTLE measure up against WiFi? a comparison of indoor locationperformance,” in European Wireless 2014; 20th European WirelessConference; Proceedings of. VDE, 2014, pp. 1–6.

[36] S. Liu, Y. Jiang, and A. Striegel, “Face-to-face proximity estimationusing bluetooth on smartphones,” IEEE Transactions on Mobile Com-puting, vol. 13, no. 4, pp. 811–823, 2014.

[37] E. S. Lohan, J. Talvitie, P. F. e Silva, H. Nurminen, S. Ali-Löytty, andR. Piché, “Received signal strength models for WLAN and BLE-basedindoor positioning in multi-floor buildings,” in 2015 InternationalConference on Location and GNSS (ICL-GNSS). IEEE, 2015, pp.1–6.

[38] J. Haverinen and A. Kemppainen, “Global indoor self-localizationbased on the ambient magnetic field,” Robotics and AutonomousSystems, vol. 57, no. 10, pp. 1028–1035, 2009.

[39] G. Berkovich, “Accurate and reliable real-time indoor positioning oncommercial smartphones,” in Indoor Positioning and Indoor Navigation(IPIN), 2014 International Conference on. IEEE, 2014, pp. 670–677.

[40] B. Gozick, K. P. Subbu, R. Dantu, and T. Maeshiro, “Magneticmaps for indoor navigation,” Instrumentation and Measurement, IEEETransactions on, vol. 60, no. 12, pp. 3883–3891, 2011.

[41] M. J. Abadi, L. Luceri, M. Hassan, C. T. Chou, and M. Nicoli, “Acollaborative approach to heading estimation for smartphone-basedPDR indoor localisation,” in Indoor Positioning and Indoor Navigation(IPIN), 2014 International Conference on. IEEE, 2014, pp. 554–563.

[42] Y. Shu, C. Bo, G. Shen, C. Zhao, L. Li, and F. Zhao, “Magicol:indoor localization using pervasive magnetic field and opportunisticWiFi sensing,” IEEE Journal on Selected Areas in Communications,vol. 33, no. 7, pp. 1443–1457, 2015.

[43] J. Torres-Sospedra, R. Montoliu, G. Mendoza, O. Belmonte, D. Ram-bla, and J. Huerta, “Providing databases for different indoor positioningtechnologies: Pros and cons of magnetic field and Wi-Fi based posi-tioning,” Mobile Information Systems, 2016.

[44] B. Li, T. Gallagher, A. G. Dempster, and C. Rizos, “How feasible isthe use of magnetic field alone for indoor positioning?” in Indoor Posi-tioning and Indoor Navigation (IPIN), 2012 International Conferenceon. IEEE, 2012, pp. 1–9.

[45] W. Storms, J. Shockley, and J. Raquet, “Magnetic field navigation inan indoor environment,” in Ubiquitous Positioning Indoor Navigationand Location Based Service (UPINLBS), 2010. IEEE, 2010, pp. 1–10.

[46] K. B. Anonsen and O. Hallingstad, “Terrain aided underwater nav-igation using point mass and particle filters,” in Proceedings of theIEEE/ION Position Location and Navigation Symposium, vol. 2006,2006, pp. 1027–1035.

[47] D. Dardari, P. Closas, and P. M. Djuric, “Indoor tracking: Theory,methods, and technologies,” Vehicular Technology, IEEE Transactionson, vol. 64, no. 4, pp. 1263–1278, 2015.

[48] M. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tutorial onparticle filters for online nonlinear/non-Gaussian Bayesian tracking,”IEEE Trans. on Signal Processing, vol. 50, no. 2, pp. 174–188, Mar.2002.

[49] M. Klepal and S. Beauregard, “A backtracking particle filter forfusing building plans with PDR displacement estimates,” in Proc. 5thWorkshop on Positioning, Navigation and Communication, Hannover,Germany, Mar. 2008, pp. 207–212.

[50] S. Beauregard, “Omnidirectional pedestrian navigation for first re-sponders,” in Proc. 4th Workshop on Positioning, Navigation andCommunication, Hannover, Germany, Mar. 2007, pp. 33–36.

[51] O. Woodman and R. Harle, “Pedestrian localisation for indoor envi-ronments,” in Proc. of UbiComp08, Seoul, Korea, Sep. 2008.

[52] B. Krach and P. Robertson, “Integration of foot-mounted inertialsensors into a Bayesian location estimation framework,” in Proc. 5th

Workshop on Positioning, Navigation and Communication, Hannover,Germany, Mar. 2008, pp. 55–61.

[53] M. Khider, S. Kaiser, P. Robertson, and M. Angermann, “The effect ofmaps-enhanced novel movement models on pedestrian navigation per-formance,” in Proc. 12th European Navigation Conference, Toulouse,France, Apr. 2008.

[54] P. Davidson, J. Collin, and J. Takala, “Map-aided autonomous pedes-trian navigation system,” in Proc. 18th Int. Conf. Integrated NavigationSystems, St. Petersburg, Russia, May 2011, pp. 314–318.

[55] C. Ascher, C. Kessler, R. Weis, and G. Trommer, “Multi-floor mapmatching in indoor environments for mobile platforms,” in Proc. of Int.Conf. on Indoor Positioning and Indoor Navigation, Sydney, Australia,Nov. 2012.

[56] S. Thrun, W. Burgard, and D. Fox, Probabilistic robotics. MIT Press,2005.

[57] H. Nurminen, A. Ristimäki, S. Ali-Löytty, and R. Piché, “Particlefilter and smoother for indoor localisation,” in Proc. Int. Conf. IndoorPositioning and Indoor Navigation, Montbéliard, France, Nov 2013.

[58] M. Khider, S. Kaiser, P. Robertson, and M. Angermann, “Maps andfloor plans enhanced 3D movement model for pedestrian navigation,”in Proc. ION GNSS, Savannah, GA, Sep. 2009.

[59] P. Gilliéron, D. Büchel, I. Spassov, and B. Merminod, “Indoor nav-igation performance analysis,” in Proc. of the European NavigationConference GNSS, 2004.

[60] M. Koivisto, H. Nurminen, S. Ali-Löytty, and R. Piché, “Graph-basedmap matching for indoor positioning,” in International Conference onInformation.

[61] S. Hilsenbeck, D. Bobkov, G. Schroth, R. Huitl, and E. Steinbach,“Graph-based data fusion of pedometer and WiFi measurements formobile indoor positioning,” in Proceedings of the 2014 ACM Inter-national Joint Conference on Pervasive and Ubiquitous Computing.ACM, 2014, pp. 147–158.

[62] I. Spassov, “Algorithms for map-aided autonomous indoor pedestrianpositioning and navigation,” Ph.D. dissertation, Ecole PolytechniqueFédérale de Lausanne (EPFL), Switzerland, 2007.

[63] P. Davidson, J. Collin, and J. Takala, “Application of particle filtersto map-matching algorithm,” J. of Gyroscopy and Navigation, vol. 2,no. 4, pp. 286–293, 2011.

[64] K. Abdulrahim, C. Hide, T. Moore, and C. Hill, “Integrating low costIMU with building heading in indoor pedestrian navigation,” J. GlobalPositioning Systems, vol. 10, no. 1, pp. 30–38, 2011.

[65] ——, “Aiding low cost inertial navigation with building heading forpedestrian navigation,” Journal of Navigation, vol. 64, no. 2, pp. 219–233, 2011.

[66] J. Borenstein, L. Ojeda, and S. Kwanmuang, “Heuristic reduction ofgyro drift for personnel tracking systems,” Journal of Navigation,vol. 62, no. 01, pp. 41–58, 2009.

[67] J. Borenstein and L. Ojeda, “Heuristic drift elimination for personneltracking systems,” Journal of Navigation, vol. 63, no. 4, pp. 591–606,2010.

[68] A. Jiménez, F. Seco, F. Zampella, J. Prieto, and J. Guevara, “Improvedheuristic drift elimination (iHDE) for pedestrian navigation in complexbuildings,” in Proc. Int. Conf. Indoor Positioning and Indoor Naviga-tion, Guimarães, Portugal, Sep. 2011.

[69] P. Davidson, M. Kirkko-Jaakkola, J. Collin, and J. Takala, “Navigationalgorithm combining building plans with autonomous sensor data,”Gyroscopy and Navigation, vol. 6, no. 3, pp. 188–196, 2015.

[70] Z. Syed, J. Georgy, C. Goodall, A. Noureldin, N. El-Sheimy, andM. Atia, Methods of Attitude and Misalignment Estimation for Con-straint Free Portable Navigation. United States Patent ApplicationPublication, Pub.No.: US2012/0245839A1, Pub. Date: Sep. 27 2012.

[71] V. Renaudin, C. Combettes, and F. Peyret, “Quaternion based headingestimation with handheld mems in indoor environments,” in Position,Location and Navigation Symposium-PLANS. IEEE/ION, 2014, pp.645–656.

[72] C. Combettes and V. Renaudin, “Comparison of misalignment estima-tion techniques between handheld device and walking directions,” inProc. Int. Conf. on Indoor Positioning and Indoor Navigation (IPIN).IEEE, 2015.

[73] R. Harle, “A survey of indoor inertial positioning systems for pedes-trians,” IEEE Communications Surveys & Tutorials, 2013.

[74] A. Brajdic and R. Harle, “Walk detection and step counting on uncon-strained smartphones,” in Proceedings of the 2013 ACM InternationalJoint Conference on Pervasive and ubiquitous computing. ACM, 2013,pp. 225–234.

23

[75] J. Käppi, J. Syrjärinne, and J. Saarinen, “MEMS-IMU based pedestriannavigator for handheld devices,” in Proc. ION GPS, Salt Lake City, UT,Sep. 2001, pp. 1369–1373.

[76] H. Leppäkoski, J. Käppi, J. Syrjärinne, and J. Takala, “Error analysisof step length estimation in pedestrian dead reckoning,” in Proc. ofION GPS, Portland, OR, Sep. 2002.

[77] H. Weinberg, “Using the ADXL202 in pedometer and personal navi-gation applications,” Analog Devices Inc., Norwood, MA, Applicationnote AN-602, 2002.

[78] J. Saarinen, “A sensor-based personal navigation system and its ap-plication for incorporating humans into a human-robot team,” Ph.D.dissertation, Helsinki University of Technology, Finland, 2009.

[79] Q. Ladetto, “On foot navigation: Continuous step calibration using bothcomplementary recursive prediction and adaptive Kalman filtering,” inProc. ION GPS, Salt Lake City, UT, Sep. 2000, pp. 1735–1740.

[80] L. Fang, P. Antsaklis, L. Montestruque, M. McMickell, M. Lemmon,Y. Sun, H. Fang, I. Koutroulis, M. Haenggi, M. Xie, and X. Xie,“Design of a wireless assisted pedestrian dead reckoning system: TheNavMote experience,” IEEE Trans. Instrum. Meas., vol. 54, no. 5, pp.2342–2358, Dec. 2005.

[81] T. Judd, “A personal dead reckoning module,” in Proc. ION GPS, 1997,pp. 47–51.

[82] H.-H. Lee, S. Choi, and M.-J. Lee, “Step detection robust against thedynamics of smartphones,” Sensors, vol. 15, no. 10, pp. 27 230–27 250,2015.

[83] A. Rai, K. K. Chintalapudi, V. N. Padmanabhan, and R. Sen, “Zee:zero-effort crowdsourcing for indoor localization,” in Proceedings ofthe 18th annual international conference on Mobile computing andnetworking. ACM, 2012, pp. 293–304.

[84] E. Munoz Diaz and A. L. Gonzalez Mendiguchia, “Step detector andstep length estimator for an inertial pocket navigation system,” in Proc.Int. Conf. on Indoor Positioning and Indoor Navigation (IPIN). IEEE,2014.

[85] M. Susi, V. Renaudin, and G. Lachapelle, “Motion mode recognitionand step detection algorithms for mobile phone users,” Sensors, vol. 13,no. 2, pp. 1539–1562, 2013.

[86] L. Bao and S. S. Intille, “Activity recognition from user-annotated ac-celeration data,” in Pervasive computing. Springer Berlin Heidelberg,2004, pp. 1–17.

[87] L. Rong, D. Zhiguo, Z. Jianzhong, and L. Ming, “Identification ofindividual walking patterns using gait acceleration,” in Proc. 1st Int.Conf. on Bioinformatics and Biomedical Engineering, ICBBE 2007.IEEE, 2007, pp. 543–546.

[88] Y. Makihara, N. T. Trung, H. Nagahara, R. Sagawa, Y. Mukaigawa, andY. Yagi, “Phase registration of a single quasi-periodic signal using selfdynamic time warping,” in Computer Vision–ACCV 2010. Springer,2011, pp. 667–678.

[89] M. Susi, D. Borio, and G. Lachapelle, “Accelerometer signal featuresand classification algorithms for positioning applications,” in Proceed-ings of the 2011 International Technical Meeting of The Institute ofNavigation, 2001, pp. 158–169.

[90] M.-S. Pan and H.-W. Lin, “A step counting algorithm for smartphoneusers: Design and implementation,” Sensors Journal, IEEE, vol. 15,no. 4, pp. 2296–2305, 2015.

[91] V. Renaudin, M. Susi, and G. Lachapelle, “Step length estimation usinghandheld inertial sensors,” Sensors, vol. 12, no. 7, pp. 8507–8525,2012.

[92] U. Steinhoff and B. Schiele, “Dead reckoning from the pocket – anexperimental study,” in Proc. Int. Conf. on Pervasive Computing andCommunications (PerCom). IEEE, 2010, pp. 162–170.

[93] N. Wang, E. Ambikairajah, S. Redmond, B. Celler, and N. Lovell,“Classification of walking patterns on inclined surfaces from ac-celerometry data,” in Proc. 16th Int. Conf. on Digital Signal Processing,July 2009, pp. 1–4.

[94] I. Bylemans, M. Weyn, and M. Klepal, “Mobile phone-based dis-placement estimation for opportunistic localisation systems,” in Proc.Int. Conf. on Mobile Ubiquitous Computing, Systems, Services andTechnologies, 2009, UBICOMM’09. IEEE, 2009, pp. 113–118.

[95] J. Jahn, U. Batzer, J. Seitz, L. Patino-Studencka, and J. Boronat, “Com-parison and evaluation of acceleration based step length estimators forhandheld devices,” in Proc. Int. Conf. on Indoor Positioning and IndoorNavigation, Zürich, Switzerland, Sep. 2010.

[96] D. Gusenbauer, C. Isert, and J. Krösche, “Self-contained indoor posi-tioning on off-the-shelf mobile devices,” in Proc. Int. Conf. on Indoorpositioning and indoor navigation (IPIN). IEEE, 2010, pp. 1–9.

[97] A. Ali and N. El-Sheimy, “Low-cost MEMS-based pedestrian naviga-tion technique for GPS-denied areas,” Journal of Sensors, 2013.

[98] P. Davidson and J. Takala, “Pedestrian navigation combining IMUmeasurements and gait models,” J. of Gyroscopy and Navigation, vol. 4,no. 2, pp. 79–84, 2013.

[99] Q. Ladetto and B. Merminod, “Digital magnetic compass and gyro-scope integration for pedestrian navigation,” in Proc. 9th Int. Conf. onIntegrated Navigation Systems, St-Petersburg, Russia, May 2002.

[100] A. Ali, H. Chang, J. Georgy, Z. Syed, and C. Goodall, “Method andapparatus for determination of misalignment between device and pedes-trian,” Jul. 24 2014, WO Patent App. PCT/CA2014/000,040. [Online].Available: https://www.google.com/patents/WO2014110672A1?cl=ja

[101] N. Mohssen, R. Momtaz, H. Aly, and M. Youssef, “It’s the humanthat matters: accurate user orientation estimation for mobile computingapplications,” in Proceedings of the 11th International Conferenceon Mobile and Ubiquitous Systems: Computing, Networking and Ser-vices. ICST (Institute for Computer Sciences, Social-Informatics andTelecommunications Engineering), 2014, pp. 70–79.

[102] K. Kunze, P. Lukowicz, K. Partridge, and B. Begole, “Which way amI facing: Inferring horizontal device orientation from an accelerometersignal,” in Wearable Computers, 2009. ISWC’09. International Sympo-sium on. IEEE, 2009, pp. 149–150.

[103] M. Chowdhary, M. Sharma, A. Kumar, S. Dayal, and M. Jain, “Methodand apparatus for determining walking direction for a pedestriandead reckoning process,” May 22 2014, US Patent App. 13/682,684.[Online]. Available: http://www.google.com/patents/US20140142885

[104] S. A. Hoseini-Tabatabaei, A. Gluhak, and R. Tafazolli, “uDirect: Anovel approach for pervasive observation of user direction with mobilephones,” in Pervasive Computing and Communications (PerCom), 2011IEEE International Conference on. IEEE, 2011, pp. 74–83.

[105] O. D. Lara and M. A. Labrador, “A survey on human activity recog-nition using wearable sensors,” Communications Surveys & Tutorials,IEEE, vol. 15, no. 3, pp. 1192–1209, 2013.

[106] M. Susi, V. Renaudin, and G. Lachapelle, “Detection of quasi-staticinstants from handheld MEMS devices,” in IPIN, 2011, pp. 1–9.

[107] Z. Yang, C. Wu, Z. Zhou, X. Zhang, X. Wang, and Y. Liu, “Mobilityincreases localizability: A survey on wireless indoor localization usinginertial sensors,” ACM Computing Surveys (CSUR), vol. 47, no. 3,p. 54, 2015.

[108] Z. Syed, J. Georgy, A. Ali, H.-W. Chang, and C. Goodall, “Show-ing smartphones the way inside: Real-time, continuous, reliable, in-door/outdoor localization,” GPS World, vol. 24, no. 3, pp. 30–35, 2013.

[109] N. El-Sheimy and C. Goodall, “Everywhere navigation: Integratedsolutions on consumer mobile devices,” Inside GNSS, Sep 2011.

a survey of selected indoor positioning methods …...reliable indoor positioning. among...

Documents